0% encontró este documento útil (0 votos)

402 vistas258 páginas

4 Inteligencia Artificial, UCO, TESIS

4 Inteligencia Artificial,UCO,TESIS

Cargado por

Giuseppe

Derechos de autor

Nos tomamos en serio los derechos de los contenidos. Si sospechas que se trata de tu contenido, reclámalo aquí.

Formatos disponibles

Descarga como PDF, TXT o lee en línea desde Scribd

0% encontró este documento útil (0 votos)

402 vistas258 páginas

4 Inteligencia Artificial, UCO, TESIS

4 Inteligencia Artificial,UCO,TESIS

Cargado por

Giuseppe

Derechos de autor

Nos tomamos en serio los derechos de los contenidos. Si sospechas que se trata de tu contenido, reclámalo aquí.

Formatos disponibles

Descarga como PDF, TXT o lee en línea desde Scribd

Está en la página 1/ 258

Universidad de Cordoba,

Espana
Doctorado en Ingeniera y Tecnologa
Departamento de Arquitectura de Computadores,

Electronica
y Tecnologa Electronica

Inteligencia Artificial:

aplicaciones a la espectrometra y dosimetra neutronicas

Tesis
Que en cumplimiento parcial de los requisitos para obtener el grado de
Y TECNOLOGIA

DOCTOR EN INGENIERIA

Presenta

Jose Manuel Ortiz Rodrguez

Directores de tesis
Dr. Hector Rene Vega-Carrillo
Dr. Jose Ignacio Benavides Bentez

Febrero de 2011
Cordoba, Espana

TITULO: Inteligencia artificial: aplicaciones a la espectrometra y dosimetra

neutrnicas
AUTOR: Jos Manuel Ortiz Rodrguez
Edita: Servicio de Publicaciones de la Universidad de Crdoba. 2011
Campus de Rabanales
Ctra. Nacional IV, Km. 396 A
14071 Crdoba
www.uco.es/publicaciones
publicaciones@uco.es

ISBN-13: 978-84-694-9096-9

TTULO DE LA TESIS: INTELIGENCIA ARTIFICIAL: APLICACIONES A LA

ESPECTROMETRA Y DOSIMETRA NEUTRNICAS

DOCTORANDO/A: JOS MANUEL ORTIZ RODRGUEZ

INFORME RAZONADO DEL/DE LOS DIRECTOR/ES DE LA TESIS

(se har mencin a la evolucin y desarrollo de la tesis, as como a trabajos y publicaciones derivados de la misma).

El trabajo desarrollado, y evidenciado a travs de las publicaciones realizadas, tiene el

comn denominador de la originalidad ya que se ha propuesto una alternativa
novedosa, basada en la tecnologa de la inteligencia artificial, para resolver un
problema mal condicionado que por definicin tiene un nmero infinito de soluciones.
Este problema es la determinacin del espectro de neutrones, E(E), as como 13
valores de tres magnitudes dosimtricas, E, Hp,s(10, ) y H*(10), a partir de 7 tasas de
conteo medidas con un sistema espectromtrico de Esferas Bonner.
En sus primeros productos el doctorando demostr la validez del uso de redes
neuronales artificiales para resolver el problema cientfico. El uso de redes neuronales
artificiales se aplica sin que exista un procedimiento para seleccionar la arquitectura de
la red que resuelva el problema, as que la primera aproximacin se hizo mediante
ensayo y error, lo que llevo mucho tiempo encontrar la topologa adecuada de la red.
Para abatir esta dificultad el doctorando procedi a buscar mtodos de optimizacin de
la topologa de la red y propuso el uso de la tcnica de optimizacin de Taguchi, con lo
que obtuvo mejores resultados, a la par del desarrollo de estos procedimientos
incursiono en el diseo de interfaces amigables para el usuario. Con estos desarrollos
el doctorando dirigi proyectos de trabajos de fin de carrera, que en Mxico les
llamamos tesis de licenciatura, cuya relevancia mereci que el trabajo recibiera
distinciones nacionales y que fueran publicados; tambin ha dirigido tesis de maestra
cuyos productos han recibido distinciones en congresos por la calidad y la relevancia
de los mismos. Con estas acciones el doctorando cumpli con otra fase de la
formacin doctoral que es la formacin de capital humano de calidad.
En la tercera fase de la evolucin del doctorando, paso a sustituir la metodologa
Taguchi de optimizacin para utilizar Algoritmos Genticos para seleccionar los pesos
sinpticos ptimos para entrenar una red neuronal, el producto de esta aproximacin
es que el uso de algoritmos genticos y las redes neuronales permitieron desarrollar
un procedimiento que permite entrenar varios centenares de redes neuronales y
obtener la mejor de ellas en un periodo de minutos. Solo para resaltar este hallazgo,
me permito indicar que en la primera fase del trabajo la red neuronal se seleccionaba
sin procedimiento y su entrenamiento llevaba varias semanas de tiempo de cmputo.
Las bondades de este nuevo mtodo de reconstruccin de espectros es que no se
requiere de un espectro inicial, como es el caso de los mtodos que se usan y que

emplean mtodos iterativos de solucin y tampoco requiere de tiempos largos de

cmputo, como los requeridos por mtodos Monte Carlo, cuando se usan en la
reconstruccin. La innovacin desarrollada por el doctorando han sido validados
mediante experimentos. El conjunto de artculos desarrollados por el doctorando en
esta lnea de investigacin lo volvieron notorio en la comunidad cientfica y fue invitado
a redactar un captulo para un libro cuyo tema central son las redes neuronales
artificiales.
Los trabajos publicados que contienen los mritos arriba mencionados son los
siguientes,
1.

Ortiz-Rodrguez, J.M.; Martnez-Blanco, M.R. and Vega-Carrillo, H.R. Evolutionary

artificial neural networks in neutron spectrometry. Chapter X, published in:
Yacamoto, M. (editor). ARTIFICIAL NEURAL NETWORKS XXX, Ed. INTECH
(2011). [ISBN: 978-953-7619-243-2]

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R.; Gallego, E. and Vega-Carrillo, H.R.

Neutron spectrometry and dosimetry based on a new approach called Genetic
Artificial Neural Networks. 12th Congress of the International Radiation Protection
Association (IRPA12). IAEA Procc. Series STI/PUB/1460. International Atomic
Energy Agency, Vienna, Austria. (2010). [ISBN: 978-92-0-105410-4]

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009). A

computational tool design for Evolutionary Artificial Neural Networks in Neutron
Spectrometry and Dosimetry. Electronics, Robotics and Automotive Mechanics
Conference 2009: CERMA 09: 113-118. [IEEE Computer Society, ISBN: 9787995-3799-3]

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Evolutionary Artificial Neural Networks in Neutron Dosimetry. Transactions of the
American Nuclear Society, 110: 647-648. Ed. American Nuclear Society,
Washington DC, USA. ISBN: 0003-018X (Print)]

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Artificial Neural Networks modeling evolved genetically, a new approach applied in
Neutron Spectrometry and Dosimetry research areas. Electronics, Robotics and
Automotive Mechanics Conference 2008: CERMA 08: 387-392. [IEEE Computer
Society, ISBN: 978-7995-3799-3]

Por lo antes expuesto, se autoriza la presentacin de la tesis doctoral.

Crdoba, _9_ de _diciembre_ de 2010

Firma del/de los director/es

Fdo.: Dr. Hctor Ren Vega - Carrillo_

Fdo.: Jos Ignacio Benavides Bentez

Inteligencia Artificial:

aplicaciones a la espectrometra y dosimetra neutronicas

c
Derechos de autor 2011
Jose Manuel Ortiz-Rodrguez

Todos los derechos reservados

Ninguna parte de e sta publicacion podra ser reproducida, almacenada o transmitida en

cualquier forma o medio electronico, mecanico,

de fotocopiado, de almacenamiento o cualquier
otra forma, sin el consentimiento previo y por escrito del autor. El permiso se debera solicitar
directamente a: Jose Manuel Ortiz-rodrguez: Telefono +52(492)923-9407 Ext. 1515/1507;
Buzon electronico: morvymm@yahoo.com.mx.

a personas o
Nota: El autor no asume responsabilidad alguna por cualquier lesion y/o dano
propiedades por la negligencia o el uso de los productos aqu presentados o cualquier otra
circunstancia. Tampoco se responsabilidad alguna por cualquier uso u operacion de
cualquiera de los metodos, productos, instrucciones o ideas contenidos en e ste material.

Mexico MMXI

TTULO DE LA TESIS: INTELIGENCIA ARTIFICIAL: APLICACIONES A LA

ESPECTROMETRA Y DOSIMETRA NEUTRNICAS

DOCTORANDO/A: JOS MANUEL ORTIZ RODRGUEZ

INFORME RAZONADO DEL/DE LOS DIRECTOR/ES DE LA TESIS

(se har mencin a la evolucin y desarrollo de la tesis, as como a trabajos y publicaciones derivados de la misma).

El trabajo desarrollado, y evidenciado a travs de las publicaciones realizadas, tiene el

emplean mtodos iterativos de solucin y tampoco requiere de tiempos largos de

Ortiz-Rodrguez, J.M.; Martnez-Blanco, M.R. and Vega-Carrillo, H.R. Evolutionary

artificial neural networks in neutron spectrometry. Captulo de libro publicado en:
Chi Leung Patrick Hui. (editor) ARTIFICIAL NEURAL NETWORKS APPLICATION, Ed. INTECH (2011). [ISBN: 978-953-307-658-4]. In Press.

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R.; Gallego, E. and Vega-Carrillo, H.R.

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009). A

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Evolutionary Artificial Neural Networks in Neutron Dosimetry. Transactions of the
American Nuclear Society, 110: 647-648. Ed. American Nuclear Society,
Washington DC, USA. ISBN: 0003-018X (Print)]

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Por lo antes expuesto, se autoriza la presentacin de la tesis doctoral.

Crdoba, _9_ de _diciembre_ de 2010

Firma del/de los director/es

Fdo.:___________________

Fdo.: __________________

Inteligencia Artificial:

aplicaciones a la espectrometra y dosimetra neutronicas

Aprobada por el jurado:

Dedicatoria

Chayis. Por el enorme sacrificio que te he demandado. Por todo el tiempo sacrificado a
y que
la convivencia familiar; Por todo el apoyo que necesitaste al dar a luz a nuestro retono

no supe ni te pude dar, estando a tu lado acompan andote

y apoyandote;
Por todo el desvelo,
preocupaciones y sacrificios extremos a que te he sometido y que sin condiciones has aceptado.

Richy. Por haberles faltado en los primeros das de tu vida y que nunca sabre como
pagar, recuperar y enmendar. Hijo, eres mi mas grande alegra y orgullo.

La familia de Chayis, mi familia, sin cuyo apoyo y soporte antes, durante y despues de
la realizacion de e ste proyecto, hubiera sido posible la visita que nos hace Richy, y con ella toda

y las palabras
la alegra y animos
de vivir que nos ha inyectado. El tiempo no me alcanzara,
y apoyo.
me faltan, para agradecer todo su carino

VII

Agradecimientos

Agradezco todo el apoyo brindado por parte del Dr. Hector Rene Vega-Carrillo, Docente Investigador de la Universidad Autonoma de Zacatecas, Mexico, quien a lo largo de mi forma
cion profesional y en la construccion de mi proyecto de vida me ha acompanado
y respaldado,

observando mi evolucion, orientandome,

aconsejandome
y dirigiendome, pero sobre todo y mas
importante, por haber trascendido y evolucionado el trato diario y cotidiano, convirtendolo en
una amistad sincera, solida e imperecedera.

Agradezco tambien por todo el tiempo que me dedico para comentar y corregir los resultados obtenidos en las distintas actividades realizadas y en los comentarios y sugerencias que
permitieron darle forma y calidad a las teoras y herramientas desarrolladas a lo largo de la
realizacion de la presente tesis doctoral.

Agradezco al Dr. Jose Ignacio Benavides Bentez, Profesor titular de la Universidad de

por todo el tiempo que ha dedicado en apoyar la realizacion de e ste proyecCordoba en Espana,

to. Desde las primeras platicas

que realizo en la Universidad Autonoma de Zacatecas, Mexico,
hasta el presente, siempre ha estado al pendiente y dispuesto a prestar la ayuda necesaria
para culminar con e xito el reto emprendido.

Agradezco al Dr. Eduardo Gallego, Profesor titular y jefe del Departamento de Ingeniera
por sus comentarios emitidos en
Nuclear de la Universidad Politecnica de Madrid en Espana,
el proceso de revision de los diversos trabajos que fueron presentados en congresos o enviados
para su publicacion en revistas. Agradezco tambien por el interes mostrado en las teoras y
herramientas tecnologicas que fueron desarrolladas a lo largo de las investigaciones llevadas
a cabo durante el doctorado.

Agradezco al Dr. Roberto Mendez, Responsable del Laboratorio de Patrones Neutronicos,

perteneciente al Laboratorio de Metrologa de Radiaciones Ionizantes del CIEMAT en Es-

VIII

por haber propiciado el hacer posible la realizacion de una estancia de investigacion

pana,
pre-doctoral en las instalaciones del centro donde labora. Agradezco tambien por el tiempo
de los codigos desarrollados para el sistema espectrometrico
dedicado a la planeacion y diseno
con el que trabaja.

Agradezco a Jose Mara de los Arcos Merino, jefe en su momento, del Laboratorio de
por el el apoyo y respaldo
Metrologa de Radiaciones Ionizantes del CIEMAT en Espana,
brindados desde mi llegada al CIEMAT. Agradezco tambien por la confianza que deposito en
mi persona, y por el tiempo que se dio para discutir y comentar las ideas y teoras que pudieron
darle forma a la cuantificacion de la incertidumbre asociada con la medicion de las tasas de
conteo de un sistema de esferas Bonner, mediante la estadstica de Poisson.

Agradezco a Jorge Guerrero Araque, con quien conviv dentro y fuera del CIEMAT, todo
el tiempo que duro mi estancia. A traves de la convivencia diaria aprend a ver la vida desde
otra perspectiva, la de su propia persona y la de su pas, y con quien compart experiencias y
como individuo.
puntos de vista que me hicieron crecer mas

Agradezco a la Agencia Espanola

de Cooperacion Internacional para el Desarrollo (AECID) por haber confiado en m y haberme seleccionado para realizar una estancia pre-doctoral

en el CIEMAT de Espana.

Agradezco a todos aquellos que no me es posible enumerar, amigos y familiares, por el

apoyo y animo
infundidos a lo largo de los altibajos vividos durante la realizacion de e ste
proyecto, sin los cuales, la culminacion de e sta meta hubiera sido imposible.

A todos ustedes: MUCHAS GRACIAS!

Resumen

El objetivo de la presente tesis doctoral se centra en el estudio de metodologas apli de arquitecturas o ptimas de redes neuronales, as como en el diseno
de
cadas en el diseno
herramientas tecnologicas que permitan utilizar el conocimiento adquirido por las redes pre

viamente disenadas,
en el ambito
de la reconstruccion de espectros de neutrones y calculo
de
dosis equivalentes en ambientes laborales, experimentales y de laboratorio. Debido a que los
neutrones no son desviados por un campo electrico o magnetico, y tambien porque no producen
una ionizacion apreciable, se deben emplear metodos indirectos para detectarlos y medir su
energa. La medicion y particularmente la determinacion de espectros de neutrones, no es una
tarea trivial. Para resolver este problema existen metodos e instrumentos que permiten determinar los espectros de neutrones, a partir de los cuales es posible calcular la dosis ambiental y
la dosis personal equivalente. El sistema espectrometrico de esferas Bonner, compuesto por un
detector de neutrones termicos colocado en el centro de un conjunto de esferas de polietileno

alta densidad, cuyos diametros

varan desde 2 hasta 18 pulgadas, es uno de los metodos mas
de la esfera hace que el sistema tenga una eficiencia de deteccion de
utilizados. El tamano
neutrones en funcion de la energa. El conjunto de eficiencias conforma una matriz de respuesta, que se utiliza junto con las tasas de conteo del detector para determinar el espectro.

y los intervalos de clase de la energa de los

Debido a que el numero
de esferas es pequeno
neutrones es grande, el problema se traduce en resolver un sistema de ecuaciones mal condi
cionado con un numero
infinito de soluciones. Para reconstruir el espectro es necesario resolver

este sistema de ecuaciones, empleandose

varios procedimientos como metodos Monte Carlo, de
parametrizacion e iterativos, mismos que presentan diversas dificultades que han motivado

el desarrollo de procesos complementarios. En las ultimas

decadas se han intentado procedimientos novedosos para el proceso de reconstruccion de espectros de neutrones, basados en las
tecnologas de Algoritmos Geneticos y Redes Neuronales Artificiales. La reconstruccion de es
pectros de neutrones y el calculo
de dosis equivalentes mediante Redes Neuronales a partir de
las tasas de conteo ledas con el Sistema de Esferas Bonner, ha recibido mayor atencion debido
a los exitosos resultados que se han obtenido, ya que resuelven muchos de los problemas exper-

de las redes no
imentados con las tecnicas clasicas.
Sin embargo, el procedimiento de diseno

es un problema trivial, ya que determinar un conjunto apropiado de valores de los parametros

estructurales y de aprendizaje de las mismas sigue siendo una tarea difcil. Asimismo, debido
a que se trata de una tecnologa emergente, se carece de herramientas tecnologicas que permitan aplicarla de forma eficiente. Las dificultades mencionadas, han motivado el desarrollo
de redes neuronales artificiales, as como el diseno
de
de estrategias para optimizar el diseno
mejorado, en comparacion
herramientas tecnologicas simples en su uso y con un desempeno
con los metodos tradicionales utilizados en la actualidad.

Abstract

The aim of the present doctoral thesis is centered in the study of methodologies for the
optimum design of neural nets architectures, as well as in the design of technological tools that
allow to use the knowledge acquired by the previously designed nets, applied in the neutron
spectra unfolding problem and equivalent doses calculus in working, experimental and laboratory atmospheres. Because the neutrons are not deviated by an electric or magnetic field,
and also because they dont produce an appreciable ionization, indirect methods should be
used to detect them and to measure their energy. The measurement and particularly the determination of neutrons spectra, is not a trivial task. To solve this problem exist methods and
instruments that allow to determine the neutrons spectra, starting from which it is possible to
calculate the environmental and equivalent personal doses. The Bonner Spheres Spectrometric System, composed by a neutorns thermal detector which is placed in the center of a group
of polyethylene spheres whose diameters varies from 2 up to 18 inches, is one of the most
methods used. The size of the sphere makes that the system has an neutrons efficiency detection in function of the energy. The group of efficiencies conforms the response matrix which
is used together with the count rates of the detector to determine the spectrum. Because the
number of spheres it is smaller than the intervals of class of the energy of the neutrons, the
problem is translated in solving an ill-conditioned equations system with an infinite number
of solutions. To unfold the neutron spectrum it is necessary to solve this equations system, being used several procedures like Monte Carlo, parametrization and iterative methods. These
methods present diverse difficulties that have motivated the development of complementary
procedures. In the last decades, novel procedures based on Artificial Intelligence techonology
such as Genetic Algorithms or Artificial Neural Networks, have been attempted in the neutron
spectrometry domain. The neutron spectra unfolding process and equivalent doses calculation
by means of Artificial Neural Networks, starting from the read count rates with the Bonner
Spheres System, has received the bigger attention due the successful results obtained. With
the use of this technology, many of the experienced problems using the classical techniques
are solved. However, the desing process of neural net architectures is not a trivial problem.

XII

To determine a suitable group of values of the learning and structural parameters of neural
nets, it continues being a difficult task. Furthermore, because it is an emergent tecnology, lack
of technological tools that allow to apply it in an efficient way is experienced. The mentioned
difficulties, have motivated the development of strategies to optimize the optimum design of
neural nets, as well as the design of technological tools simple and easy in their use and with
an improved acting, in comparison with the traditional methods used at the present time.

Indice
general
Portada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Derechos de Autor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Aprobacion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Dedicatoria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Agradecimientos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . VII
Resumen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Abstract

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1. Introduccion

1.1. Justificacion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2. Pregunta cientfica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3. Hipotesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.4. Objetivo General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

de literatura
2. Revision

2.1. Inteligencia Artificial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.1. Redes Neuronales Artificiales . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.2. Algoritmos Evolutivos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2. Espectrometra de Neutrones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2.1. Espectrometra de Neutrones con el SEEB . . . . . . . . . . . . . . . . . .

3. Materiales y metodos

3.1. Introduccion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.2. Pre-procesamiento de la informacion . . . . . . . . . . . . . . . . . . . . . . . . . .

3.3. Tecnicas de optimizacion de RNA . . . . . . . . . . . . . . . . . . . . . . . . . . . .

XIII

INDICE
GENERAL

XIV

Robusto de Redes Neuronales Artificiales . . . . . . . . . . . . . . .

3.3.1. Diseno

3.3.2. Redes Evolucionadas por Algoritmos Evolutivos . . . . . . . . . . . . . . .

de herramientas tecnologicas . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4. Diseno

3.4.1. VMCNPex y generacion automatica

de matrices de entrenamiento de RNA 72
3.4.2. NSDann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.4.3. NSDEann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.4.4. NSDTB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.5. Optimizacion de las herramientas disenadas

utilizando la estadstica de Poisson

3.6. Aplicacion de las herramientas en experimentos reales y comparacion con tecni

cas clasicas

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4. Resultados
4.1. Revistas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

89
96
96

4.2. Captulo de libro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

5. Conclusiones

197

Indice
de figuras
2.1. Taxonoma de la Inteligencia Artificial . . . . . . . . . . . . . . . . . . . . . . . . .

2.2. Esquema basico

de una neurona biologica . . . . . . . . . . . . . . . . . . . . . . .

2.3. De la neurona biologica a la neurona artificial . . . . . . . . . . . . . . . . . . . .

2.4. modelo de una neurona artificial . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.5. Esquema de una Red neuronal de una capa con S neuronas . . . . . . . . . . . . .

2.6. Esquema general de una RNA de 3 capas . . . . . . . . . . . . . . . . . . . . . . .

2.7. Posible clasificacion de las redes neuronales . . . . . . . . . . . . . . . . . . . . . .

2.8. Entrenamiento de una Red Neuronal . . . . . . . . . . . . . . . . . . . . . . . . . .

2.9. Estructura celular biologica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.10.Analogas en los algoritmos geneticos I . . . . . . . . . . . . . . . . . . . . . . . . .

2.11.Analogas en los algoritmos geneticos II . . . . . . . . . . . . . . . . . . . . . . . .

2.12.Analogas en los algoritmos geneticos III . . . . . . . . . . . . . . . . . . . . . . . .

2.13.Modelo de fitness que ilustra la forma de operar de un AE . . . . . . . . . . . . .

2.14.Convergencia entre RNA y AG . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.15.Estructura evolutiva de un algoritmo genetico . . . . . . . . . . . . . . . . . . . .

2.16.Efectos intensificadores y diversificadores de los operadores cruce y mutacion . .

2.17.Poder de penetracion de partculas radiactivas . . . . . . . . . . . . . . . . . . . .

2.18.Deteccion de neutrones empleando medios moderadores

. . . . . . . . . . . . . .

2.19.Sistema Espectrometrico de Esferas Bonner . . . . . . . . . . . . . . . . . . . . . .

2.20.Combinacion detector-moderador . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.21.Tecnicas clasicas
para espectrometra de neutrones con el SEEB

. . . . . . . . .

2.22.Espectrometra de neutrones con el SEEB y tecnicas de IA . . . . . . . . . . . . .

3.1. Esquema general de RNA aplicadas en la espectrometra neutronica . . . . . . .

INDICE
DE FIGURAS

XVI

3.2. Tecnicas de Optimizacion de RNA: DRRNA y RNAE . . . . . . . . . . . . . . . . .

3.3. Codigos para espectrometra de neutrones basadas en RNA . . . . . . . . . . . .

3.4. Codigos para espectrometra de neutrones basadas en RNA para un detector de

3 He

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.5. Comparacion de metodos clasicos

y los basados en IA en espectrometra neutronica 51
3.6. Metodologa utilizada para la obtencion de los datos del entrenamiento de la RNA. 52
3.7. Ejecucion del codigo MCNP en el entorno MS-DOS de Windows . . . . . . . . . .

3.8. Codigo que automatiza la creacion de matrices de entrenamiento de RNA en

espectrometra Neutronica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.9. Entorno grafico

de VMCNPex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

robusto de RNA . . . . . . . . . . . . . . . . . . . . . . . . .
3.10.Metodologa de diseno

de RNAE . . . . . . . . . . . . . . . . . . . . . . . .
3.11.Metodologa general de diseno

de RNA por medio de AE . . . . . . . . . . . . . . . . . . .

3.12.Metodologa de diseno

3.13.Ventana principal de NGO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.14.Desarrollo de Herramientas de computo en base a MDRRNA y RNAE . . . . . .

de herramientas . . . . . . . . . . . . . . . . . . . . . . . . .
3.15.Estrategia de diseno

3.16.Codigo NSDann6LiI(Eu) para espectrometra neutronica con RNA . . . . . . . .

3.17.Archivo de tasas de conteo medidas con el SEEB para ingresar a NSDann . . . .

3.18.Seleccion del archivo de entrada con las lecturas BSS al codigo NSDann . . . . .

3.19.Codigo NSDann3He para espectrometra neutronica con RNA . . . . . . . . . . .

3.20.Codigo de reconstruccion de espectros de neutrones NSDEann6LiI . . . . . . . .

3.21.Archivo de tasas de conteo sin normalizar para NSDEann . . . . . . . . . . . . .

3.22.Herramienta de prediccion de NGO para espectrometra neutronica . . . . . . .

normalizados a uno . . .
3.23.Archivo con espectros reconstruidos con NGO que estan

3.24.Codigo de reconstruccion de espectros de neutrones NSDEann3He . . . . . . . .

3.25.Necesidad de herramienta para comparar diversas metodologas . . . . . . . . .

3.26.NSDTB:compara espectros obtenidos con diversas metodologas . . . . . . . . . .

3.27.Archivo con espectros y dosis a comparar con el codigo NSDTB6Li . . . . . . . . .

3.28.Distribucion de Poisson en tasas de conteo ledas con el SEEB . . . . . . . . . . .

3.29.Espectros y dosis calculados con NSDann, alrededor de la distribucion de Poisson 88

3.30.Espectros y dosis con incertidumbres, calculadas con la distribucion de Poisson .

INDICE
DE FIGURAS

XVII

3.31.SEEB con detector 6 LiI (Eu) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.32.SEEB con detector 3 He . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.33.Electronica asociada del SEEB para medir las tasas de conteo . . . . . . . . . . .

3.34.Ambiente de codigos tradicionales para espectrometra neutronica . . . . . . . .

3.35.Ambiente de trabajo del codigo UMG . . . . . . . . . . . . . . . . . . . . . . . . . .

3.36.Espectros y dosis calculados con el codigo NSDAnn y sus incertidumbres . . . . .

5.1. Espectrometra de neutrones con el SEEB y tecnicas de IA . . . . . . . . . . . . . 198

5.2. Contribucion de tecnicas de IA en espectrometra de neutrones con el SEEB . . . 198

Indice
de cuadros
2.1. Clasificacion de los neutrones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.1. Variables de diseno

y sus niveles . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.2. Variables de ruido y sus niveles . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.3. Arreglo Ortogonal cruzado L9 (34 ) - L4 (22 ) . . . . . . . . . . . . . . . . . . . . . . .

3.4. Parametros

de diseno
con la metodologa RNAE . . . . . . . . . . . . . . . . . . .

XVIII

Captulo 1

Introduccion
amplio la Inteligencia Artificial (IA) [14], indica la capacidad que
En su sentido mas
tiene un artefacto para realizar los mismos tipos de funciones que caracterizan la capacidad

de aprendizaje del humano [5,6]. La busqueda

de la IA ha tomado dos caminos fundamentales:
uno de los caminos es la investigacion psicologica y fisiologica de la naturaleza del aprendizaje

humano [79], y el otro camino es el desarrollo tecnologico de sistemas informaticos

cada vez
complejos [1015]. En este sentido, la IA se ha aplicado a sistemas y programas computamas
cionales capaces de realizar tareas complejas simulando el proceso de aprendizaje humano,
se encuentra considerablemente lejos de este [1619].
aunque aun

En la actualidad existen diversas tendencias para el desarrollo de sistemas basados en

la tecnologa de IA, tales como: Sistemas Expertos (SE) [20], Logica Difusa (LD) [21, 22], Algoritmos Evolutivos (AE) [2328], Redes Neuronales Artificiales (RNA) [2937], entre otros.
Los SE tratan de reproducir el razonamiento humano de forma simbolica, las RNA y los AE
biologico. Las RNA recrean la estructura del cerebro
lo hacen desde un punto de vista mas
humano imitando el proceso de aprendizaje [38, 39] y, los AE utilizan la filosofa Darwininana
apto en la busqueda

de la supervivencia del mas

de una solucion o ptima [4049].

El objetivo de la IA es el desarrollo de paradigmas que requieren las maquinas

para desarrollar tareas cognoscitivas en las que los humanos normalmente son mejores [1, 23, 50, 51].
Un sistema de IA debe ser capaz de hacer tres cosas: almacenar conocimiento, aplicar el
conocimiento almacenado en la solucion de problemas y adquirir nuevo conocimiento por
medio de la experiencia [52, 53]. Un sistema de este tipo, tiene tres componentes clave: representacion, razonamiento y aprendizaje. En esta esfera, los campos de investigacion en que se
1

1. Introduccion

han aplicado las tecnicas de IA son: procesamiento de informacion, reconocimiento de modelos,

en el area

juegos y areas
aplicadas como el diagnostico medico, y durante los ultimos
anos
de
las Ciencias Nucleares. [54, 55]

La teora de RNA ha brindado una alternativa a la computacion clasica

para aquellos
problemas en los que los metodos tradicionales han entregado resultados no muy convincentes
exitosas de las RNAs son: procesamieno poco convenientes [5661]. Las aplicaciones mas

to de imagenes
y de voz, reconocimiento de patrones, planeacion, interfaces adaptivas para

sistemas hombre/maquina
(robotica), prediccion, control y optimizacion, filtrado de senales
y

recientemente en la reconstruccion de espectros de neutrones y calculo

simultaneo
de dosis

equivalentes por medio del Sistema Espectrometrico de Esferas Bonner (SEEB) en el area
de
las Ciencias Nucleares. [6265]

Para una correcta evaluacion de la dosis efectiva recibida por los trabajadores en difer
entes instalaciones nucleares, se requieren dosmetros, personales o de area,
calibrados con
fuentes de neutrones cuyos espectros son similares a aquellos encontrados en esas instalaciones [6672]. Sin embargo, se encuentran muchos problemas al intentar determinar el espectro de energa de neutrones de un dato medido. El conocimiento de estos espectros de neutrones se puede obtener por medio de tecnicas espectrometricas [67, 73], asumiendo en todos
linealmente relacionados al espectro que se busca deterlos casos que los datos medidos estan
minar a traves de la ecuacion integro-diferencial de Fredholm de primer tipo. En general, el
espectro de neutrones deseado se obtiene invirtiendo numericamente la ecuacion, utilizando
un codigo de reconstruccion que define implcitamente una solucion de la misma. [74]

Las tecnicas utilizadas para la deteccion de los neutrones, se describen ampliamente

en numerosas referencias [6670, 72, 73, 7588]. Entre todas las tecnicas disponibles para la
utilizado para los propositos de la protecespectrometra de neutrones, el SEEB es el mas

cion radiologica [8993], debido a las multiples

ventajas que ofrece el tener la informacion
delicada de la espectrometra basada en el SEEB, es el proceso de
espectral. La parte mas
reconstruccion, para esto se han desarrollado varios codigos de deconvolucion computarizados
basados en este sistema, cuyos puntos crticos son en general: la complejidad de los codigos, la

1. Introduccion

necesidad de un usuario experto y de proporcionar informacion real a priori [94].

De acuerdo a lo anterior, durante los ultimos

anos,
se han realizado intentos por desarrollar nuevos codigos de deconvolucion de espectros de neutrones como: MITOM [95], FRUIT [74],
MAXED y GRAVEL [96], que se basan en algoritmos de reconstruccion iterativos, cuyo funcionamiento requiere de un espectro inicial, a partir del cual se realizan las iteraciones. El
problema de esta situacion es que el espectro que resulta del proceso de reconstruccion depende de que tan semejante sea el espectro inicial del espectro que se esta buscando. El otro
para
problema asociado a estos codigos es su pobre disponibilidad y que su manejo no es facil
el usuario.

utilizando nuevos metodos basados en tecnologas de IA para

En la actualidad se estan
realizar la espectrometra y dosimetra de neutrones, utilizando principalmente la teora de

recibiendo gran atencion por parte de los investiRNA [97104]. Estas

tecnicas de IA estan
gadores, debido a los resultados exitosos obtenidos, sin embargo, en el actual estado del arte

que tiene que ver con el uso de la IA en el ambito

de la espectrometra de neutrones, aun

existen problemas con lo que se refiere a la seleccion o ptima de parametros

de RNA y AG. Ac
tualmente los investigadores determinan e stos parametros
por medio de la tecnica de ensayo
y error.

Lo anterior implica que no es facil

para el disenador
elegir una topologa de algoritmo
o red satisfactorios, sin importar la experiencia que tenga. Asimismo, debido a que es una
tecnologa emergente, la carencia de herramientas cientficas y tecnologicas que estimulen

su aplicacion practica
en situaciones reales, como en el caso del presente proyecto en la es
pectrometra neutronica, representa un obstaculo
para su desarrollo. Por consiguiente, debe
sencilla y eficiente de superar e stas desventajas.
existir una forma mas

1.1.

Justificacion
Los instrumentos de radioproteccion operacionales utilizados en la actualidad, tienen re-

spuestas dependientes del espectro neutronico. Todo ello hace difcil la interpretacion de las
indicaciones de estos instrumentos, si no se conoce la distribucion espectral de fluencia de los

1. Introduccion

neutrones en los puntos de medida [6670, 72, 73, 7588].

que como el espectro puede cambiar de manera apreciable dentro de una

Resulta ademas
instalacion nuclear, la espectrometra de neutrones es de interes capital para la comprension

y correcta calibracion de los instrumentos usados para medir la dosis, ya que unicamente
este
metodo permite una evaluacion precisa de las magnitudes dosimetricas de interes y al mismo
tiempo elaborar los protocolos y estrategias adecuados para la proteccion radiologica [8993].

utilizando nuevos metodos basados en tecnologas de IA para reAl presente, se estan

alizar la espectrometra y dosimetra de neutrones, utilizando principalmente la teora de
recientes el uso de la tecnologa de redes neuronales artificiales en la
RNA [6265]. En anos
espectrometra y dosimetra neutronicas se ha convertido en un procedimiento alternativo con
el que se han obtenido resultados satisfactorios [97104]. Sin embargo, a pesar de sus ventajas,
de las redes no es un problema trivial, ya que determinar un
el procedimiento mismo de diseno

conjunto apropiado de valores de los parametros

estructurales y de aprendizaje de las mismas
sigue siendo una tarea compleja.

La estructura de la red, tambien conocida como la topologa de red, as como los parametros de aprendizaje de una red neuronal, son factores que contribuyen de manera significativa
de la misma, sin embargo, se ha observado que los investigadores en esta
con el desempeno

area
de conocimiento realizan la seleccion de los parametros
de la red a traves de la tecnica
de ensayo y error, empleando en este proceso largos periodos de tiempo, obteniendo redes neu y baja capacidad de generalizacion.
ronales de pobre desempeno

Las dificultades intrnsecas asociadas con las RNA y los AG hacen difcil aplicar las

teoras de IA en el ambito
de la espectrometra neutronica, lo que ha motivado el estudio y
de RNA y AG. Aunado a lo anterior, debido a que se trata
desarrollo de metodologas de diseno
de una tecnologa emergente, se observa que se carece de herramientas cientfico-tecnologicas

que permitan aplicar e sta tecnologa de forma agil

y eficiente, en ambientes laborales, experimentales y de laboratorio.

1. Introduccion

Se ha observado que los investigadores que han aplicado la tecnologa de RNA en la

espectrometra de neutrones carecen de herramientas que les permitan llevar a cabo con efi de
ciencia y rapidez las siguientes actividades: Crear matrices de entrenamiento para el diseno

distintas topologas de red; aplicar alguna estrategia metodologica y sistematicas

para identi
una vez entrenadas las redes, aplicar el conocimiento
ficar los parametros
o ptimos de diseno;
adquirido en ambientes experimentales reales a traves de herramientas tecnologicas creadas

a la medida y, comparar los resultados obtenidos con las redes disenadas

contra los obtenidos

con herramientas tradicionales de forma rapida

y efectiva.

La reconstruccion de espectros de neutrones y el calculo

de dosis equivalentes a partir
de las tasas de conteo provenientes del SEEB, empleando tecnicas de IA como son las RNA
y AG, es un procedimiento alternativo que ha estado recibiendo gran atencion durante los

debido a los exitosos resultados obtenidos. Sin embargo, en vista de lo novedoso

ultimos
anos

del metodo utilizado, se observa que los investigadores no utilizan una estrategia sistematica

de las RNA y que demas,

se carece de
para la seleccion de los parametros
o ptimos de diseno
herramientas tecnologicas que permitan aplicar de forma eficiente el conocimiento adquirido

por las redes disenadas

en ambientes experimentales reales, simplificando con ello los procedimientos y tiempos de respuesta para la obtencion del espectro y la dosis deseados.

A pesar de que la tecnologa de la Inteligencia Artificial ofrece muchas ventajas en la

reconstruccion de espectros y la determinacion de la dosis utilizando el SEEB, existen varios problemas que deben ser resueltos antes de que esta tecnologa se consolide en la espectrometra y la dosimetra de neutrones. Tambien es necesario desarrollar herramientas tecnologicas que permitan utilizar la tecnologa de la Inteligencia Artificial con el fin de que el
usuario final las adopte y las aplique.

En vista lo expuesto con anterioridad, se formula la siguiente pregunta cientfica.

1.2.

Pregunta cientfica

Cuales
son las estrategias necesarias que deben utilizarse para disenar
una

RNA capaz de reconstruir el espectro de neutrones y determinar la dosis utilizando

1. Introduccion

las tasas de conteo de un SEEB?

como unica
pieza de informacion

Para responder a la pregunta anterior, se plantea la siguiente hipotesis.

1.3.

Hipotesis
El uso de la Metodologa de Diseno
Robusto de RNA y los Algoritmos Evolutivos en el

diseno
de una RNA permitira que e sta sea capaz de reconstruir el espectro de neutrones y determinar la dosis utilizando como unica

pieza de informacion las tasas de conteo de un SEEB.

Con la intencion de confirmar la validez de la hipotesis cientfica de esta investigacion,

el objetivo general es el siguiente.

1.4.

Objetivo General
o ptimo de arquitecturas RNA y disenar
herramienExplorar estrategias para el diseno

tas tecnologicas en el ambito

de la espectrometra y dosimetra de neutrones.

Para alcanzar el objetivo general, se proponen los siguientes objetivos particulares:

de arquitecEstudiar, analizar e implementar tecnicas de optimizacion para el diseno
turas de redes neuronales artificiales.
analizadas en el punto anterior, en la optimizacion de
Aplicar las estrategias de diseno,

RNA en el ambito
de la espectrometra neutronica.
herramientas para simplificar el proceso de creacion de las matrices de entreDisenar

namiento de las las RNA a ser disenadas.

un codigo para la espectrometra de neutrones que permita aplicar en ambientes
Disenar

experimentales reales, el conocimiento adquirido por las RNA disenadas.

de las RNA disenadas

Disenar
de herramientas que permitan analizar el desempeno

y comparar los resultados contra las tecnicas clasicas

utilizadas en espectrometra de
neutrones.

Captulo 2

de literatura
Revision
2.1.

Inteligencia Artificial
amplio la Inteligencia Artificial (IA) [1, 23, 50, 51], indica la capacidad
En su sentido mas

que tiene un artefacto para realizar los mismos tipos de funciones que caracterizan al pensamiento humano [2, 5]. El termino IA se define como la habilidad para aprender de forma
efectiva, reaccionar de forma adaptiva, realizar decisiones apropiadas, comunicar y enten
der ideas, a traves del lenguaje o las imagenes,
de forma sofisticada [3, 4, 12]. La IA busca
un razonamiento logico exacto en base a la manipulacion de smbolos y comprende metodos,
herramientas y sistemas para resolver problemas que normalmente requieren la inteligencia
humana [7, 17].

Un sistema de IA tiene tres componentes clave: representacion, razonamiento y aprendizaje [8, 9, 19, 20], por lo que un sistema de IA debe ser capaz de hacer tres cosas: almacenar
conocimiento, aplicar el conocimiento almacenado en la solucion de problemas y adquirir nuevo conocimiento por medio de la experiencia. La posibilidad de desarrollar un artefacto as ha

despertado la curiosidad del ser humano desde la antiguedad;

sin embargo, no fue hasta la
segunda mitad del siglo XX cuando esa posibilidad se materializo en herramientas tangibles.

Desde entonces, los cientficos han tratado de implementar en maquinas

y computadoras la
inteligencia humana de varias formas [13, 14].

La IA clasica
nacio poco despues de la maquina
Von Newmann, durante la misma era

intelectual [15]. El termino Inteligencia Artificial fue acunado

en 1956 por John McCarthy, del
se celebro la conferencia de Dartmouth,
Instituto Tecnologico de Massachussets. En ese ano
7

de literatura
2. Revision

en Hanover (Estados Unidos), y en ella, McCarthy, Marvin Minsky, Nathaniel Rochester y

Claude E. Shannon establecieron las bases de la IA como un campo independiente dentro de

la informatica.

En la IA clasica,
el e nfasis se centra en la construccion de representaciones simbolicas, las cuales se llaman as porque presumiblemente representan algo [1315]. Siguiendo un

lenguaje de conceptos como modelo para la IA clasica,

se encuentra que las representaciones
stica [1619]. Al igual que las expresiones del
simbolicas poseen una estructura casi lingu

lenguaje natural, las expresiones de la IA clasica

generalmente son complejas, construidas de

forma sistematica
a traves de smbolos simples. Dada una existencia limitada de smbolos, se
podran componer nuevas expresiones con significado en virtud de la composicionalidad de las

expresiones simbolicas y la analoga entre la estructura sintactica

y la semantica.

Desde el punto de vista del conocimiento, la IA asume la existencia de representaciones

mentales, y modela el conocimiento como un procesamiento secuencial de representaciones
simbolicas [8, 9, 19, 20]. Probablemente la inspiracion del procesamiento secuencial viene de
la naturaleza secuencial del lenguaje natural y la inferencia logica, como muchas de las es
tructuras de la maquina
Von Newmann. Este procesamiento es secuencial como en la programacion tpica de computadoras, ya que aun cuando no hay un orden predeterminado, las
operaciones se desarrollan de la forma paso a paso.

Los objetivos principales de la IA son: desarrollar metodos y sistemas para resolver problemas, generalmente resueltos por la actividad intelectual de los humanos [1, 23, 50, 51], por

ejemplo, el reconocimiento de imagenes,

el procesamiento del lenguaje y del habla, planeacion

y prediccion, mejorando con ello la informacion de los sistemas informaticos

[1015]; el segundo objetivo fundamental de la IA, es desarrollar modelos que simulen a los organismos vivos
y al cerebro humano en particular, mejorando con ello nuestro entendimiento acerca de como
trabaja el cerebro humano [79].

Con el avance de la ciencia moderna, la busqueda

de la IA ha tomado dos caminos fundamentales: la investigacion psicologica y fisiologica de la naturaleza del pensamiento hu-

de literatura
2. Revision

mano [7, 9, 17, 50], y consecuentemente el desarrollo tecnologico de sistemas informaticos

cada
complejos. En este sentido, el termino IA se ha aplicado a sistemas y programas invez mas

formaticos
capaces de realizar tareas complejas, simulando el funcionamiento del pensamiento humano, aunque todava muy lejos de alcanzar este objetivo. En esta esfera los campos
importantes son el procesamiento de la informacion, el reconocimiento
de investigacion mas

de modelos, los juegos y las areas

aplicadas como el diagnostico medico o la espectrometra y
dosimetra de neutrones [54, 55].

Como se muestra en la figura 2.1, actualmente existen diversas tendencias en cuanto al

desarrollo de sistemas de Inteligencia Artificial [11]: Redes Neuronales Artificiales (RNA) [29
36, 5660, 105], Sistemas Expertos (SE) [20], Logica Difusa (LD) [22], Algoritmos Evolutivos
(AE) [2428, 53], etc.

Figura 2.1: Taxonoma de la Inteligencia Artificial

Los sistemas expertos intentan reproducir el razonamiento humano de forma simboli bica. Los algoritmos geneticos y las redes neuronales lo hacen desde una perspectiva mas

de literatura
2. Revision

inspirados en la naturaleza, sin embargo, las redes

ologica [106, 107]. Ambos modelos estan
orientadas al aprendizaje de un solo individuo, recreando la estructura de
neuronales estan
un cerebro humano [38], lo cual se define como aprendizaje fenotpico, mientras que los algoritmos evolutivos se interesan en la adaptacion de una poblacion a un entorno cambiante,
esto se define como aprendizaje genotpico [4049]. A pesar de la complejidad de los sistemas
descritos, los resultados distan mucho de un autentico pensamiento inteligente, ya que los sistemas simbolicos de la IA tienen muy poco poder al tratar con informacion inexacta, incierta,
corrupta, imprecisa o ambigua.

Los sistemas simbolicos de la IA se han asociado en las ultimas

decadas con dos aspectos principales: afrontar los problemas de la representacion y el procesamiento (razonamien
to) [5, 9]. Aunque la IA utiliza un riguroso razonamiento logico matematico
[35, 108], este no
es flexible y es difcil de implementar debido principalmente a que los metodos de IA son parcialmente exitosos en la implementacion de la inteligencia humana, pero que sin embargo,
esta muy lejos de e sta. Actualmente, las principales direcciones en el desarrollo de la IA
aun
son el desarrollo de metodos y sistemas para resolver problemas por medio de la IA, de forma
distinta a como la realiza el humano, pero proporcionado resultados similares, por ejemplo, los
sistemas expertos; y desarrollar metodos y sistemas para resolver problemas por medio de la
IA modelando el pensamiento humano en la forma en la que trabaja fsicamente el cerebro,
por ejemplo, las redes neuronales artificiales.

y el
Las redes neuronales artificiales ofrecen un paradigma atractivo para el diseno

analisis
de sistemas adaptativos inteligentes para un amplio rango de aplicaciones en IA
por muchas razones, entre las cuales se puede incluir a las siguientes: flexibilidad para la
adaptacion y el aprendizaje (mediante la modificacion de las estructuras computacionales empleadas), robustez ante la presencia de ruido (datos erroneos o incompletos), habilidad para
generalizar, capacidad de recuperacion ante fallas, potencial para su computacion masiva paralela, y semejanza (si bien superficial) con las redes neuronales biologicas

de literatura
2. Revision

2.1.1.

Redes Neuronales Artificiales

importantes trabajos sobre neurologa fueron realizados por SanLos primeros y mas
tiago Ramon y Cajal (1852-1934) y Charles Scott Sherrington (1852-1957). A partir de sus

estudios se conoce que el elemento basico

que conforma el sistema nervioso son las neuronas.
La morfologa de una neurona biologica individual se representa de manera simplificada en
la figura 2.2, en donde pueden distinguirse tres partes fundamentales: el soma o cuerpo de la
celula, las dendritas y el cilindro-eje o axon [33, 58, 59, 109, 110].

Figura 2.2: Esquema basico

de una neurona biologica

Las dendritas son fibras encargadas de recibir las senales

electricas procedentes de otras

neuronas o tejidos organicos

y transmitirlas al soma. Las multiples
senales
procedentes de las
dendritas son procesadas por el soma y transmitidas al axon. El cilindro-eje o axon es una fibra
de gran longitud (en relacion con el resto de la neurona), conectada al soma por un extremo y
del
que se divide en el otro en una serie de ramificaciones nerviosas. El axon recoge la senal

soma y la transmite a otras neuronas o tejidos organicos

(como por ejemplo los musculos).

Para hacer una transicion a partir de las ideas provenientes de la neurobiologa hacia las
estructuras idealizadas que forman la base de la mayora de los modelos de RNA, en primer

de literatura
2. Revision

lugar se debe describir el modelo general de la neurona artificial [111]. En la figura 2.3 se
observa un elemento de procesamiento en forma general (neurona artificial) y su similitud con
una neurona biologica.

Figura 2.3: De la neurona biologica a la neurona artificial

Si se realiza una comparacion entre una neurona biologica y una neurona artificial, se
encuentran las siguientes analogas: al igual que una neurona biologica, la neurona artificial
tiene muchas entradas pero solo una salida que se puede conectar a muchas otras neuronas en
la red. La i-esima neurona recibe las entradas de la j-esima neurona, indicado como X j , notese
que este valor es tambien la salida de la j-esima neurona, ya que la salida generada por la
i-esima neurona esta etiquetada como X j . Cada conexion a la i-esima neurona tiene asociado
con ella una cantidad llamada peso sinaptico.

El peso en la conexion de la j-esima neurona a

la i-esima neurona se describe por Wi j [29, 30, 33, 34, 58, 59, 111].

Cada neurona determina un valor de entrada neto en base a todas las conexiones de
entrada. En la ausencia de conexiones especiales, la entrada de la red se calcula tpicamente
sumando los valores de entrada multiplicados por sus correspondientes pesos. En otras palabras, la entrada de red a la i-esima neurona se puede escribir como se muestra en la ecuacion
2.1, donde el ndice j se ejecuta sobre todas las conexiones de la neurona.:

redi =

X jWi j

(2.1)

j =1

Una vez que se calcula la entrada de la red, e sta se convierte en un valor de activacion
para la neurona. Este valor de activacion se puede escribir como se indica en la ecuacion 2.2

de literatura
2. Revision
para denotar que la activacion es una funcion explcita de la entrada de la red.

ai (t ) = Fi (ai (t 1), redi (t ))

(2.2)

Notese que la activacion actual podra depender del valor previo de la activacion, a(t 1).
En la mayora de los casos, la activacion y la entrada de la red son identicos, y regularmente
los terminos se utilizan de forma intercambiable.

Una vez que se calcula la activacion de la neurona, se puede determinar el valor de salida
aplicando una funcion de salida:

Xi = fi (ai )

(2.3)

Debido a que generalmente, a = redi , esta funcion se escribe normalmente como:

Xi = fi (redi )

(2.4)

Aunque una sola neurona puede realizar ciertas funciones simples de procesamiento de

informacion, un unico
nodo o neurona no es suficiente para muchos problemas practicos,
por

lo que regularmente se utilizan redes con un gran numero

de neuronas. Debido a lo anterior

sera util
pensar en una red de neuronas y en la coleccion de valores de los pesos sinapticos

como un sistema dinamico

[31, 60, 105].

El desarrollo de las Redes Neuronales Artificiales (RNA) comenzo con los estudios de McCulloch y Pitts en la decada de 1940, con la creacion de la primera red neuronal [32, 38, 112].
conocidos, el Perceptron, fue desarrollado por Frank Rosenblatt
Uno de los tipos de redes mas
a final de 1950 [37, 61, 113]. Tras el impulso inicial, las investigaciones sobre el tema se detuvieron repentinamente en 1969, cuando Minsky y Papert demostraron la incapacidad del
Perceptron para resolver determinados tipos de problemas (problemas XOR). Los estudios se
retomaron posteriormente a principios de 1980, con la aparicion de nuevos tipos de RNA con
capacidades para la optimizacion no lineal.

de literatura
2. Revision

Las RNA son modelos muy simplificados de los procesos realizados por el cerebro, y estan

inspiradas en el modelo biologico de la neurona. Una RNA esta compuesta de un gran numero
interconectados y
de e stos elementos simples de procesamiento (neuronas), los cuales estan
operan en paralelo. Cada neurona se conecta a las otras neuronas por medio de enlaces de

comunicacion directa llamados pesos sinapticos,

mismos que en conjunto con las neuronas
constituyen la estructura neuronal. Los pesos representan la informacion utilizada por la red
para resolver un problema dado.

La Figura 2.4, muestra el modelo de neurona artificial. Aqu, la entrada P, un vector de

R elementos de entrada, se representa por la barra de color gris a la izquierda de la figura.
Las dimensiones de P se muestran como Rx1. Luego de que ingresan a la red e stas entradas, se

multiplican por la matriz de pesos sinapticos

W . En la figura puede observarse que a la neurona ingresa una constante con valor 1 y se multiplica por el valor de ganancia b, obteniendose
como salida el valor n. La informacion anterior, esto es la suma de b + W p, ingresa a una
funcion de transferencia f , produciendose la salida a.

Figura 2.4: modelo de una neurona artificial

Las RNA suelen estar formadas por varias neuronas interconectadas. La disposicion y
forma de conexion vara de unos tipos de redes a otras, pero de forma general las neuronas se
disponen en capas. Se puede crear una capa de neuronas que tengan diferentes funciones de
transferencia poniendolas simplemente en paralelo. Se debe resaltar que todas las neuronas
deberan tener las mismas entradas y cada neurona debera crear una salida, sin embargo,

de literatura
2. Revision

al crear una capa de neuronas, no existe restriccion alguna con relacion a que el numero
de
entradas y salidas deba ser el mismo.

En una red neuronal, la capa de entrada no es en s una capa de neuronas sino que son

a la red. En esta capa no se procesa

los puntos donde se recogen las senales
que se entregaran
la informacion. Las capas ocultas recogen la informacion de la capa de entrada y la procesan.

El numero
de capas ocultas, el numero
de neuronas por capa y la forma en que se conectan
varan de unas redes a otras, siendo este uno de los principales problemas que enfrenta el

disenador
de redes neuronales. La capa de salida recibe la informacion de la ultima
capa oculta y la transmite al medio.

En la figura 2.5 se presenta una red de una sola capa con S neuronas, en la cual cada una
de las R entradas se conecta a cada una de las neuronas, la matriz de pesos tiene ahora S filas.
La capa incluye la matriz de pesos W, los sumadores , el vector de ganancias n, la funcion de
transferencia f y el vector de salida a [114].

Figura 2.5: Esquema de una Red neuronal de una capa con S neuronas
poderosas que las redes de una sola capa, por ejemplo, una
Las redes multicapa son mas
red de dos capas que tenga una funcion sigmoidal en la primera capa y una funcion lineal en
la segunda, se puede entrenar para aproximar muchas funciones de forma aceptable, una red

de literatura
2. Revision

de una sola capa no podra hacer esto [37]. Si se considera una red con varias capas, o red
multicapa, cada capa tendra su propia matriz de pesos W, su propio vector de ganancias b,
de una capa
un vector de entradas netas n y un vector de salidas a. En el caso de existir mas
oculta, como la mostrada en la figura 2.6 donde se aprecia el esquema de una red de tres capas,
las salidas de las neuronas de la primera capa oculta son las entradas de las de la siguiente
hasta alcanzar la salida.

Figura 2.6: Esquema general de una RNA de 3 capas

su topologa, forma
En general las RNAs se pueden clasificar de diversas maneras, segun
de aprendizaje (supervisado o no supervisado), tipos de funciones de activacion y valores de
entrada (binarios o continuos). Una posible clasificacion se muestra en la figura 2.7.

La clasificacion presentada en esta figura emplea como primer criterio el tipo de entrada
que se presenta a la red y, posteriormente, si el aprendizaje es supervisado o no, es decir, si
es necesario o no conocer las salidas convenientes a las entradas empleadas durante el entrenamiento de la red. Cabe mencionar sin embargo, que existen otros tipos de redes menos
extendidas y que no se muestran en la figura 2.7.

de literatura
2. Revision

Figura 2.7: Posible clasificacion de las redes neuronales

El arreglo de neuronas en capas y los patrones de conexion dentro y entre las capas se
conoce como arquitectura de red [31, 32, 37, 38, 60, 61, 105, 112, 113]. De acuerdo con la ausencia o presencia de conexiones de realimentacion en una red, se puede distinguir dos tipos de
arquitecturas: arquitectura hacia delante y arquitectura con realimentacion. La idea central

del las RNA, en donde w y b son parametros

ajustables de la neurona, es que tales parametros
se puedan ajustar por medio del aprendizaje o entrenamiento, de tal forma que la red exhi comportamiento deseado. En funcion del tipo de red este proceso puede diferir mucho.
ba algun

La habilidad de aprender de una RNA se realiza al aplicar un algoritmo de aprendizaje

o entrenamiento. Los algoritmos de entrenamiento se clasifican en tres grupos principales:
Supervisado, No supervisado y aprendizaje reforzado [31, 32, 37, 38, 60, 61, 105, 112, 113]. El
utilizado debido a que se utilizan dos
algoritmo de entrenamiento supervisado ha sido el mas
conjuntos de datos para el entrenamiento, un vector de datos de entrada p y un vector de datos
de salida t con los valores deseados. El entrenamiento se realiza hasta que la RNA aprende
a asociar cada entrada del vector p con su correspondiente valor de salida deseado del vector
t. Desde un punto de vista muy general y de acuerdo con la ecuacion 2.5, donde ti es la salida
esperada para la entrada pi , la modificacion de los pesos de una RNA en cada iteracion puede
hacerse sumando al vector de pesos Wi , el vector de entradas pi multiplicado por el error (i )
obtenido para dicha entrada.

de literatura
2. Revision

Wi+1 = Wi + i pi = Wi + (ti ai ) pi

(2.5)

En el caso de las ganancias, la expresion empleada habitualmente es la que se muestra

en la ecuacion 2.6

bi = bi + 1 +

(2.6)

En la figura 2.8, se observa el entrenamiento de una RNA con un tipo de red de aprendizaje supervisado, es decir, para poder adaptar su comportamiento al problema planteado se
le debe ofrecer el resultado esperado a una serie determinada de entradas. Las salidas ofreci el error
das por la red a estas entradas son comparadas con las salidas esperadas y se evalua
cometido. En funcion de ese error se modifican los pesos de los enlaces y los valores de las
ganancias de las neuronas. Este proceso se repite hasta conseguir que el error alcance una
cota superior determinada, punto en el cual se considera que la red esta entrenada.

Figura 2.8: Entrenamiento de una Red Neuronal

A pesar del aparente e xito de las RNA, existen algunos aspectos que no garantizan el
uso universal de las mismas en la solucion de problemas en el mundo real. Uno de los mayores
inconvenientes en el uso de la tecnologa de RNA en la actualidad es el problema de encontrar
una estructura (arquitectura) apropiada para una tarea dada, principalmente debido a que es

de literatura
2. Revision

y estructura de la red que se necesita para resolver

muy difcil conocer de antemano el tamano
un problema dado [115120]. Una estructura ideal es una estructura que independientemente

de la iniciailizacion de los pesos sinapticos

de la red, siempre aprende la tarea, esto es, casi no
existe error en la comparacion del conjunto de datos producidos por la red cuando se comparan
generaliza bien.
con los datos deseados y ademas

Las RNA son bien conocidas por su efectividad en la representacion de sistemas de proce
sos no lineales. El poder de calculo
de una RNA proviene de de la interconexion de las neuronas

en la red, y la forma en la que se conectan las mismas determina el proceder del calculo.
Lo
de la red. El problema con
anterior constituye una importante decision en la etapa de diseno
se tiene que seleccionar un conjunto de parametros

las RNA es que para su diseno

antes de
que pueda realizarse el entrenamiento. Sin embargo, en la actualidad no existen reglas claras

para seleccionar tales parametros,

los cuales determinan el e xito en el entrenamiento. Entre
las limitaciones de las RNA, se deberan resaltar las siguientes:
La arquitectura de la red. Actualmente existe una falta de reglas fijas o guas sis
o ptimo de RNA. Debido a que no existe un conocimiento previo
tematicas
para el diseno
acerca de la complejidad del problema, la arquitectura de la red se establece tpicamente
de forma arbitraria, esto es, se realiza a traves de la tecnica de ensayo y error. Lo ante
de la red
rior orilla a la red a desempenare
con incertidumbre debido a que el tamano
de la misma: redes muy pequenas
no son capaces de aprender
influencia el desempeno
bien, pero redes demasiado grandes podran conducir al sobre ajuste. Entonces, se nece
o ptimo de RNA que puedan
sitan algoritmos o estrategias sistematicas
para el diseno
encontrar de manera apropiada la arquitectura de RNA aplicadas en cualquier problema particular. Lo anterior incluye la determinacion de el numero o ptimo de neuronas en
cada capa as como el numero de capas necesarias.

Algoritmos de entrenamiento. Actualmente no existe un algoritmo unico

de entrenamiento que se pueda emplear de manera generalizada para la solucion de problemas
por medio de RNA. Durante la fase de entrenamiento de las RNA y de acuerdo con el
algoritmo de entrenamiento utilizado, pueden quedar atrapadas en un mnimo local del
espacio global de soluciones. Lo anterior no es muy malo si el mnimo local estuviera
muy cerca del mnimo global, pero es situaciones reales, no existe garanta de que este

de literatura
2. Revision

sea siempre el caso. Aunque el algoritmo de entrenamiento de retro propagacion (BP, de

sus siglas en ingles) se ha utilizado ampliamente, su uso no garantiza el encontrar una
solucion o ptima global. El entrenamiento podra resultar en un modelo de RNA que sea

exacto unicamente
en la zona de operacion de los datos con los que la red fue entrenada
la seleccion de algunos parametros

pero que sea inexacto en otras zonas. Ademas,

en el
algoritmo de entrenamiento de propagacion inversa, tales como el momento o la tasa de

aprendizaje, tambien carecen de una gua sistematica

para su seleccion.
Datos de entrenamiento. La calidad y cantidad de los datos de entrenamiento es un
aspecto importante para el modelado de RNA. Generalmente, el e xito de una RNA recae
fuertemente en una gran cantidad de datos, lo que demanda mas tiempo de computo en
el entrenamiento. Para reducir la cantidad de datos manteniendo la calidad del modelo,
los datos utilizados deben ser cuidadosamente seleccionados para asegurar que son lo
suficiente ricos. Esto demanda un entendimiento del proceso que se desea resolver. Adicionalmente, para eliminar ruido y datos atpicos, los datos procesados podran requerir
un procesamiento previo antes de aplicarlos en el desarrollo del modelo de la RNA.
Conjunto de datos de entrenamiento. Debido a que normalmente es imposible pre
sentarle a la red todas las entrada posibles, unicamente
se le presenta una parte de e stas,
lo que se conoce como conjunto de datos de entrenamiento. Este conjunto se debe elegir de
tal forma que la red tambien proporcione una salida correcta para cualquier entrada no
incluida en el conjunto de datos de entrenamiento. Si la red responde bien a las entradas
no incluidas en los datos de entrenamiento, se dice que generaliza bien. La generalizacion

es muy similar a la interpolacion en matematicas.

A menudo una RNA se entrena con
un conjunto de patrones (el conjunto de datos de entrenamiento) y se prueba con otro (el
conjunto de datos de prueba). Si el conjunto de datos de entrenamiento no es una buena

representacion de todas las entradas posibles, la red probablemente no se desempenara

bien con las entradas que no fueron incluidas en el conjunto de entrenamiento.
de pesos sinapticos.

inicializacion
Otro problema asociado con las RNA es el lugar

en el que se comienza la busqueda

del error en la superficie de errores, a traves de la

seleccion inicial de los pesos sinapticos,

lo que se realiza de forma aleatoria. Lo anterior
determina en gran medida si se va a encontrar o no, una buena solucion. En el caso de

de literatura
2. Revision

que se encontrara una buena solucion que se desempenara

bien en el conjunto de datos
mal en el conjunto general de entradas, si
de entrenamiento, la red se podra desempenar
el conjunto de datos de entrenamiento no fuera representativo del entorno del problema
la red que este siendo disenada.

donde se va a desempenar
interferencia. Esto ocurre cuando se supone que una red aprende tareas similares al

mismo tiempo. Aparte del hecho de que redes pequenas

no son capaces de aprender
muchas asociaciones, se saturan despues de cierta cantidad de asociaciones aprendidas,
por lo que existe el peligro de que sea difcil de separar los patrones de entrada y que la
red no encuentre la manera de hacerlo.
de procesos. Siendo un metodo de caja negra para el modelado, las RNA no
Relacion
son capaces de explicar y analizar las relaciones entre las entradas y las salidas. Esto
podra ocasionar dificultades en la interpretacion de los resultados obtenidos con la red.
Todas estas limitaciones han motivado a los investigadores a generar ideas para crear o

en el proceso de
combinar a las RNA con otros metodos en la busqueda
de un mejor desempeno
de las mismas. Algunos de los esquemas que se exploran en la actualidad tienen que ver
diseno
con sistemas expertos, metodos estadstico [115120], logica difusa [21, 22, 29, 49, 121], transformacion de ondoletas as como con metodos neuro-evolutivos [42, 48, 51, 107, 122125]. En el
presente proyecto de investigacion, se exploran los metodos estadsticos y neuro-evolutivos en
de arquitecturas o ptimas de RNA.
el proceso de diseno

Una neurona artificial es una abstraccion matematica

del funcionamiento de las neuronas biologicas. La evolucion a traves de la seleccion natural es una idea central en biologa,
y el concepto de la evolucion natural ha influenciado en gran medida nuestro punto de vista
de los sistemas biologicos. De igual forma, la evolucion de los sistemas artificiales es un componente importante de la vida artificial, proporcionando una importante herramienta de mod automatico.

elado y un metodo de diseno

Los Algoritmos Evolutivos (AE) son actualmente los
modelos mas prominentes y mas ampliamente utilizados de evolucion en los sistemas de vida

artificial, mismos que se han utilizado como herramientas para resolver problemas practicos
y
como modelos cientficos de procesos evolutivos. La interseccion entre los AE y la vida artificial
incluye ambos.

de literatura
2. Revision

2.1.2.

Algoritmos Evolutivos

La resolucion de un problema complejo determinado puede asimilarse a la busqueda

de
una solucion o ptima (o al menos sub-optima) en un espacio de soluciones extenso [2428, 53].

La naturaleza se enfrenta a un problema de este estilo en su busqueda

de la mejor adaptacion
de los individuos al medio. Los miembros de una determinada poblacion compiten entre ellos

en la busqueda
de la supervivencia [43, 126].

Aquellos individuos capaces de adaptarse mejor a las condiciones del medio que les rodea
mayor probabilidad de sobrevivir [40, 127, 128]. Los miembros de la poblacion con
tendran

mayor habilidad para conseguir alimentos, soportar las condiciones climaticas

o defenderse
de una esperanza de vida mayor, de una mayor probabilidad de
de otras especies, disfrutaran
una descendencia mas
numerosa. De esta manera, las caractersticas
reproducirse y lograran

del individuo que han determinado su mayor habilidad para desenvolverse en el medio seran
transmitidas a su descendencia. Por otra parte, los individuos menos dotados y adaptados al
una probabilidad menor de trasmitir dichas caractersticas a las siguientes
entorno tendran
generaciones.

La informacion genetica determina el poder de adaptacion de cada especie (y dentro de

adaptados sobreviven y se reproducen en
cada especie, de cada individuo). Los individuos mas
mayor medida dando lugar al proceso de cambio dirigido denominado Seleccion Natural. Dicho
proceso resulta evidentemente eficaz y, sobre todo, robusto, dado que permite la adaptacion exitosa de las poblaciones en ambientes mutables e inciertos. Los AE pertenecen a una clase de

algoritmos de busqueda
estocastica
basada en poblaciones, inspirados a partir de la evolucion
natural, mismos que se basan en el principio de la supervivencia del mas apto, as como en el
fenomeno natural de la herencia genetica y la teora Darwiniana de la supervivencia.

Antes de profundizar en la exposicion del presente tema, se proporcionan algunas definiciones de los terminos empleados, a fin de comprender los conceptos a ser expuestos:
Espacio de soluciones: Conjunto de todas las posibles soluciones a un problema determinado que es posible alcanzar con el sistema de resolucion empleado. Equivale a espacio
de individuos.

de literatura
2. Revision

o alternativa: Configuracion compatible con las restricciones del problema y

Solucion
que le da solucion.
Individuo: Equivalente analogicamente a solucion o alternativa.
Conjunto de individuos existentes en un momento (iteracion) dado.
Poblacion:
Proceso de creacion de nuevos individuos. Tambien se emplea como sinoniGeneracion:
mo de poblacion.
objetivo o funcion
de evaluacion:
Funcion empleada para evaluar la caliFuncion
dad de las soluciones. Analogicamente medira la adaptacion de los individuos a su en
torno, denominandose
entones adaptabilidad. Generalmente es una funcion definida sobre el espacio de soluciones S y con imagen en R:
FO = S R
Los problemas planteados pretenden encontrar individuos en S que minimicen (o en su
caso maximicen) su imagen en R.
optima:

Solucion
Solucion s S tal que FO(s) sea o ptimo. Analogicamente, el individuo
mejor adaptado a su entorno.
sub-optima: Solucion de calidad cercana a la de la solucion o ptima o bien, de
Solucion
calidad aceptable para las condiciones del problema planteado.
Secuencia de elementos que representa (codifica)
Cadena, estructura o codificacion:
a una solucion o individuo. En la analoga natural es un cromosoma.
Gen: Analoga natural de cada uno de los elementos que conforman la cadena o cromosoma que representa un individuo.
Alelo: Valor que puede adoptar un gen.
Genotipo: Se empleara para denotar el contenido genetico de un individuo, es decir, el
cromosoma que lo codifica.
Medio externo: Entorno en el que se desarrollan y compiten los individuos. En el caso

del presente trabajo sera analogo

al dominio de ubicacion de las actividades.

de literatura
2. Revision

Fenotipo: Caractersticas fsicas de un individuo determinadas por su genotipo y las

condiciones del medio externo.
A continuacion se presentan superficialmente los mecanismos de la herencia y la termi
nologa utilizada en esta area,
con el objetivo es evidenciar la analoga entre los algoritmos
geneticos y aquello que pretenden simular.

Cada individuo de cada una de las especies que habitan en nuestro planeta poseen ciertas caractersticas que lo identifican. Estas caractersticas constituyen lo que se denomina el
fenotipo de un individuo, el cual es resultado de la interaccion del medio ambiente en que se
desarrolla y la herencia que este recibe de sus ancestros. El genotipo de un individuo designa
la constitucion genetica completa (expresa o latente) de un individuo, es decir, el genotipo es
la informacion contenida en los alelos de los genes del individuo. El fenotipo de cada individuo
esta determinado por las protenas que produce, y esto a su vez esta definido en la informacion

genetica de cada una de sus celulas. La informacion acerca de cuales

protenas se produciran
esta contenida en los cromosomas del individuo [128].

El conjunto de rasgos o caractersticas observables del individuo (fenotipo) que han determinado su e xito o su fracaso, son la expresion de la informacion contenida en su carga
genetica, [45, 46]. Dicha informacion se encuentra codificada en los genes, una secuencia de
nucleotidos (fragmento de ADN) en una determinada localizacion de un cromosoma especfico.

Los cromosomas, generalmente agrupados por parejas, se encuentran en el interior del nucleo

celular, como se aprecia en la figura 2.9. En cada celula somatica

(aquellas que constituyen
el organismo) existen dos juegos de cromosomas que definen las mismas caractersticas; un
juego es aportacion del padre del individuo y el otro lo es de la madre.

En terminos sencillos, un cromosoma es una larga molecula de ADN (Acido

Desoxirri simples llamados bases o nucleotibonucleico), formada por cuatro distintos compuestos mas
dos: adenina (A), guanina (G), citosina (C) y timina (T). Cada subcadena de tres nucleoti
dos codifica un aminoacido
diferente que al unirse con los generados por otros tercetos, formara una protena. A las subcadenas de tres nucleotidos se les llama codones y al conjunto de
nucleotidos que codifican una protena completa se les llama genes. El valor que posee un gen

de literatura
2. Revision

determinado se denomina alelo.

Figura 2.9: Estructura celular biologica

Durante la reproduccion, cada uno de los padres transmite a su descendencia parte de

caracsu carga genetica. Como resultado de este cruce los individuos descendientes poseeran

tersticas feneticas de ambos progenitores. Dado que los individuos mejor dotados poseeran
de la combinacion de las bueuna mayor descendencia, las sucesivas generaciones disfrutaran
nas caractersticas de generaciones pasadas, lo que se traducira en una mejor adaptacion al
medio.

Es frecuente que el codigo de los elementos del dominio del problema utilice un alfabeto
binario (0s y 1s) [129131]. Una vez que se ha definido la manera de codificar los elementos
del dominio del problema y se conoce la forma de pasar de un elemento a su codigo y viceversa,
es necesario fijar un punto de partida. Los algoritmos geneticos manipulan conjuntos de codigos en generaciones sucesivas. Nuevamente haciendo una analoga, manipulan poblaciones
de una vez. El algoritmo se encargara de
de codigos. En e stas un codigo puede aparecer mas
favorecer la aparicion en la poblacion de codigos que correspondan a elementos del dominio
que esten proximos a resolver el problema [25, 45, 46]. En resumen, el algoritmo recibira como entrada una poblacion de codigos y a partir de e sta generara nuevas poblaciones, donde
mientras que otros, que se mapean en mejores soluciones posialgunos codigos desapareceran
frecuencia hasta que se encuentra una satisfactoria o hasta que se
bles, aparecen con mas

de literatura
2. Revision

cumple alguna otra condicion de terminacion.

La figura 2.10 representa la analoga entre una solucion a un problema y su codificacion,

y un individuo y su genotipo.

Figura 2.10: Analogas en los algoritmos geneticos I

La relacion entre la codificacion de una solucion y la propia solucion estara influida por
el dominio de ubicacion de las actividades, de la misma forma que el fenotipo de un individuo
del medio en el que se desenvuelve, como se aprecia en la
depende de su genotipo y, ademas,
lugar a individuos
figura 2.11 (a). Dado un medio externo concreto, diferentes genotipos daran
diferentes. De la misma forma, dado un dominio de ubicacion fijo, como el mostrado en la
lugar a diferentes soluciones [47].
figura 2.11 (b), diferentes codificaciones daran

de literatura
2. Revision

(a) Una codificacion

(b) Varias codificaciones

Figura 2.11: Analogas en los algoritmos geneticos II

En la figura 2.12, puede apreciarse que ocurre en el caso de que una misma codificacion
se exponga a diferentes dominios.

Figura 2.12: Analogas en los algoritmos geneticos III

Observando la figura 2.12 podra pensarse que una determinada cadena da lugar a individuos diferentes en funcion del entorno a la que se expone, con lo que se incumplira la
condicion anterior. Sin embargo, lo que ocurre en realidad es que un determinado genotipo
expuesto a diferentes medios, produce diferentes expresiones fsicas del mismo individuo, es

de literatura
2. Revision

decir, diferentes fenotipos de un mismo individuo. Una de las condiciones para implementar
correctamente un algoritmo genetico, es que exista una correspondencia biunvoca entre el espacio de soluciones y el espacio de las codificaciones. Esto es, cada codificacion representa una
y solo una solucion, y toda solucion puede ser codificada una y solo de una manera.

Se dice que los AE son metodos de busqueda

ciega, debido a que e stos no disponen de
informacion del problema a resolver que aquella obtenida de la funcion objetivo empleamas

da. Esta caracterstica junto con el hecho de que la busqueda

se realiza simultaneamente
sobre
o menos amplio, hacen de los algoritmos geneticos una tecnica
un conjunto poblacional mas
robusta de resolucion de problemas. A esto contribuye el denominado paralelismo implcito de
los algoritmos geneticos; debido al tipo de codificacion empleado (generalmente binario), del
procesamiento de cada individuo (del codigo de cada individuo) se obtiene informacion sobre
todos los individuos cuyos codigos siguen patrones similares.

Debido a que los AE muestrean de forma simultanea

muchos puntos en el espacio de

busqueda,
son menos susceptibles que los metodos de solucion unica,
como los basados en

RNA, a quedar atrapados en la vecindad de un mnimo local y son capaces de localizar rapi
damente regiones que valgan la pena en espacios de busqueda
de grandes dimensiones. La
figura 2.13, muestra un modelo ideal para ilustrar la forma en la que trabaja un AE. El fitness
de cada individuo en la poblacion se representa por su posicion en el modelo.

Figura 2.13: Modelo de fitness que ilustra la forma de operar de un AE

de literatura
2. Revision

En la figura 2.13, se observa que en un metodo basado en solucion unica,

como por ejemp
lo las RNA, si el punto de busqueda
inicial, mostrado en el crculo amarillo, falla al identificar

la region mas cercana al o ptimo global, el algoritmo podra quedar atrapado, ya que unica tiene que realizar la siguiente inferencia para
mente cuenta con informacion local con la cual

mejorar la solucion. Entonces, ascendera el gradiente hacia el maximo

local.

En los AE, aunque algunos individuos, mostrados en los crculos rojos, podran encon maximo

trarse cerca de algun

local, es menos probable que quede atrapado, debido a que
la poblacion en general proporciona informacion global acerca del modelo. Existe una mayor
individuo que este cerca del o ptimo global, con lo que los operoportunidad de que algun
moverse a la poblacion del AE en grandes saltos para enfocar
adores geneticos le permitiran

la busqueda
en las regiones mas fructferas del modelo.

que son estrategias generales de

Aunque los AE y las RNA tienen como elemento comun

busqueda,
ambas varan en su rango. Los AE realizan una busqueda
mas global que las RNA.
La figura 2.14 muestra la convergencia de ambas estrategias.

Figura 2.14: Convergencia entre RNA y AG

Como se aprecia en la figura 2.14, a las RNA les toma mas tiempo alcanzar la vecindad
de una solucion o ptima, sin embargo, la alcanzan con mayor precision. Por otro lado, los AE

investigan el espacio entero de busqueda,

por lo que son capaces de alcanzar con mayor rapidez la region de las soluciones o ptimas, sin embargo, experimentan dificultades para localizar
un punto exacto. Esto sucede, debido a que al final el ajuste fino de la solucion recae casi por
completo en el operador de mutacion. El combinar ambas estrategias, RNA-AE, parece ser

de literatura
2. Revision

el mejor proceder [42, 48, 51, 107, 122125]. Los metodos NeuroEvolutivos se desempenan
tan

bien como lo hacen de forma independiente los AE o las RNA, en la busqueda

de una solucion
satisfactoria.

En la figura 2.15 se observa que el proceso evolutivo de una AG comienza con la creacion
aleatoria de una poblacion definida por el problema que se desea resolver [27, 40, 44, 47, 128,

132,133]. El algoritmo comienza creando un numero

determinado de cadenas que codifican un
conjunto de soluciones al problema objeto de estudio. Este conjunto de individuos representa
es un parametro

la poblacion inicial y su tamano

fundamental del proceso.

El criterio para valorar la calidad de cada una de las soluciones (en la analoga el grado de adecuacion de cada individuo al entorno) es el valor de la funcion objetivo. As en un
proceso denominado genericamente evaluacion, se determina el valor de la funcion para cada individuo. Una vez conocida la valoracion de cada individuo se somete a la poblacion a
se selecciona un grupo de individuos que
un proceso denominado seleccion, mediante el cual
la nueva generacion. La seleccion puede llevarse a cabo mediante diversos prodeterminaran
cedimientos, pero en general se realiza otorgando una mayor probabilidad de ser elegido a los
individuos mejor valorados.

Aunque de esta fase del algoritmo existen multiples

variantes, en general, los individuos
los individuos
elegidos forman una poblacion intermedia o auxiliar, de la que se seleccionaran
de manera efectiva. Estos seran
sometidos a los operadores geneticos (haque se reproduciran
obtener una descendencia que recoja
bitualmente cruzamiento y mutacion), que permitiran
las mejores caractersticas de los progenitores e introduzcan diversidad en la poblacion. Del

conjunto de la poblacion auxiliar y de la descendencia se escogera el numero

suficiente de individuos como para formar una nueva poblacion. Este proceso de sustitucion de los individuos de
la generacion anterior por nuevos individuos (que pueden ser identicos a los ya existentes) se
denomina reemplazo. Todo el proceso (evaluacion, seleccion, aplicacion de operadores y reemplazo) se repite hasta que se cumple un criterio de parada preestablecido.

de literatura
2. Revision

Figura 2.15: Estructura evolutiva de un algoritmo genetico

Ante toda esta diversidad se hace necesario establecer la condiciones para una correcta implementacion de los algoritmos geneticos. Para implementar de forma adecuada y con
garantas un algoritmo genetico, es necesario observar el cumplimiento de una serie de condiciones, que pueden resumirse en la correcta definicion de los siguientes aspectos:

de las soluciones. La codificacion de las soluciones en un algoritmo

Codificacion
genetico debe realizarse de forma binaria. Goldberg enuncio dos principios fundamentales que
debe cumplir una codificacion: Principio de los bloques constructivos con significado y Principio del alfabeto de smbolos mnimo. Se considera que una codificacion representa de manera
adecuada un problema cuando posee estas cinco propiedades: Completitud, coherencia, uniformidad, localidad y sencillez.

La codificacion binaria no resulta adecuada para la mayora de problemas de busqueda,

puesto que no siempre es posible encontrar una codificacion que de significacion a cada uno

de los elementos que forma la cadena. En la practica,

lograr una codificacion binaria perfecta

de literatura
2. Revision

que posea las cinco caractersticas anteriores resulta imposible. Por otra parte, en caso de que
se lograra dicha codificacion binaria perfecta, sera muy difcil que cumpliera el principio de
los bloques constructivos con significado. Este es el caso de gran parte de los problemas de
ingeniera en los que se requiere una elevada precision numerica. La representacion binaria
que requieren grandes capacidades
puede, en estos casos, originar individuos de gran tamano
de computo y almacenamiento, y que no permitan la aparicion de esquemas significativos,
cortos y de bajo orden. Por otra parte, individuos representados por cadenas extremadamente
del espacio de posibles soluciones, lo que hace disminuir el
largas, hacen aumentar el tamano
rendimiento de los algoritmos geneticos.

La codificacion no-binaria parece ofrecer resultados, que en muchos casos, superan los
obtenidos mediante codificaciones binarias. Sin embargo, la explicacion a este fenomeno no
es sencilla. Para explicar esta circunstancia, Goldberg presento una teora para tratar de explicar porque representaciones de alta cardinalidad ofrecen tan buenos resultados [Goldberg,
91]. Esta teora, llamada teora de los alfabetos virtuales, explica que el conjunto de alelos
que son empleados en la codificacion de los individuos converge en las primeras iteraciones
pequeno,
reduciendo el numero

a un conjunto mas
de alelos efectivos y disminuyendo as la
cardinalidad del conjunto.

de la poblacion
inicial. De forma general esta poblacion inicial se crea de
Creacion
manera aleatoria, aunque en algunos casos se realiza mediante el uso de tecnicas heursti variada posible, conteniendo individuos de
cas. Es conveniente que esta poblacion sea lo mas

caractersticas muy diferentes. Por otra parte, debe tratarse que el numero
de cadenas que
reducido posible.
representan individuos no factibles sea lo mas

En problemas muy restringidos puede resultar difcil localizar una cantidad suficiente de
Por
individuos factibles, como para completar una poblacion inicial de determinado tamano.
pequenos
de poblacion suele concentrar la busqueda

otra parte, emplear tamanos

del algorit
mo en zonas pequenas
del espacio de soluciones, y hace disminuir la diversidad. Este prob de poblacion variable. En este tipo de estrategias
lema puede resolverse empleando tamanos
de la poblacion va aumentando o disminuyendo en funcion del grado de converel tamano

de literatura
2. Revision

pequeno
para la poblacion inicial
gencia de la poblacion. As, es posible emplear un tamano
El
que facilite localizar suficientes individuos factibles y, posteriormente aumentar el tamano.
de poblacion inconstante es empleado tambien en ciertas variantes del los algoritmos
tamano
geneticos que limitan el tiempo de supervivencia de los individuos introduciendo el concepto
de edad del mismo.

La velocidad de convergencia del algoritmo puede depender en gran medida de la poblacion

inicial escogida. Una poblacion inicial poco variada puede dar lugar a fenomenos de deriva
genetica, y una en la que un individuo es especialmente bueno (superindividuo) puede provocar la convergencia prematura hacia o ptimos locales.

de evaluacion
y Funcion
de aptitud. Para poder evolucionar hacia poblaFuncion
adaptadas al entorno es necesario conocer el grado de aptitud de los
ciones cada vez mas
individuos, es decir, tiene que poderse evaluar cada individuo y ser comparado con el resto de
la poblacion. El mecanismo de seleccion posterior dependera de la medicion de la aptitud de
los individuos. En este sentido cabe distinguir entre funcion de evaluacion y funcion de aptitud.

El resultado de la funcion de evaluacion es lo que se persigue maximizar (o minimizar).

Es en s el objeto del problema planteado. Sin embargo, en ocasiones no es posible o conveniente emplear dicha funcion para evaluar los individuos en un algoritmo genetico.

Puede no ser sencillo expresar de manera numerica el objetivo perseguido en la resolucion del problema, y en caso de conseguirlo, la expresion de la funcion puede ser excesivamente

compleja y difcil de implementar informaticamente.

Entre los calculos
realizados por una computadora para la ejecucion del algoritmo, aquellos destinados a evaluar los individuos son, con
tiempo de proceso requieren. As pues, debe tratarse de simplificar en
diferencia, los que mas
lo posible la expresion de la funcion de evaluacion. Es interesante destacar, que los algoritmos
geneticos obtienen buenos resultados con funciones de aptitud que son simplificaciones de las
de evaluacion. Es decir, son suficientemente insensibles al error como para permitir utilizar
expresiones simplificadas mediante, por ejemplo, la linealizacion.

de literatura
2. Revision

Inconvenientes anadidos
al empleo de una funcion de evaluacion pura en el algoritmo,
son la necesidad de que e sta deba devolver siempre valores positivos y que, en algunos casos,
se debe minimizar el valor obtenido en lugar de maximizarlo. La necesidad de que los valores
devueltos por la funcion sean positivos, es debida a que e sta es una de las hipotesis de partida
sobre las que se asienta el teorema de Holland. Por lo tanto, se somete a la funcion de evalu habituales
acion a una serie de transformaciones que facilitan su uso en el algoritmo. Las mas
son el desplazamiento y el escalado.

Operadores geneticos.
Para que en la naturaleza las poblaciones evolucionen hacia
una mayor adaptacion a su entorno, los individuos de cada generacion deben ser sustituidos
por otros, de tal forma que recojan aquellas caractersticas exitosas de sus progenitores. No
obstante, esto no es suficiente si se pretende mejorar la adaptacion de la especie. Se hace necesario que los nuevos individuos aporten caractersticas feneticas novedosas, obtenidas de la
combinacion de la carga genetica de sus progenitores, o de la aparicion de cambios aleatorios
en su codificacion genetica. Esta tarea es asumida en los algoritmos geneticos por los operadores geneticos.

Existe una gran diversidad de operadores geneticos, resultantes en la mayor parte de

adaptaciones al problema particular que se trata de resolver. Sin embargo, los tres operadores

basicos
caractersticos de los algoritmos geneticos son: la seleccion, el cruce y la mutacion. Ca un papel fundamental en la mejora de la aptitud de las poblaciones;
da uno de ellos desempena

el operador seleccion es analogo

a los procesos competitivos para lograr la superviviencia y el

apareamiento; el cruce es analogo

a la combinacion de las cargas geneticas de los individuos
que se aparean y reproducen; y la mutacion simula las alteraciones que por diversas causas
se producen en el codigo genetico de algunos individuos.

El cruce es el operador por excelencia de los algoritmos geneticos. Mediante este operador se consigue el intercambio de informacion genetica entre individuos, mecanismo que
combinado con la accion del operador seleccion, provoca la transmision y combinacion en las
nuevas generaciones, de aquellos genotipos que se han expresado en fenotipos exitosos. A pri una aptitud mayor que
ori es imposible saber si las nuevas estructuras generadas poseeran

de literatura
2. Revision

sus progenitores, pero la accion combinada de la seleccion y el cruce hace posible a la larga,
identificar los bloques constructivos o grupos de genes que suponen caractersticas deseables
en los individuos. El operador cruce produce la convergencia hacia las mejores combinaciones
de la carga genetica ya presentes en la poblacion.

Los algoritmos geneticos son capaces de explorar y explotar el espacio de soluciones.

Por explotar se entiende realizar una busqueda

exhaustiva en una zona restringida del espa como un intensificador de la
cio. Esta tarea se realiza mediante el operador cruce, que actua

busqueda
en la zona del espacio de soluciones actualmente ocupada por la poblacion.

El efecto exploratorio, por el que se prospectan zonas del espacio alejadas de las soluciones presentes en la poblacion, es asumido por el operador mutacion. Como se representa

graficamente
en la figura 2.16, mientras la poblacion converge hacia un o ptimo local, esa zona
del espacio de soluciones es estudiada, posteriormente, el operador mutacion hace posible el
salto hacia otros lugares inexplorados.

Figura 2.16: Efectos intensificadores y diversificadores de los operadores cruce y mutacion

de literatura
2. Revision

La mutacion, salvo en algunas variantes, suele estar presente en casi todas las implementaciones de los algoritmos geneticos. Es frecuente considerarlo un operador secundario
frente al cruce, introducido para asegurar que todos los alelos posibles entran alguna vez en
la poblacion. Este operador introduce la variabilidad genetica aleatoria que contrarresta el
efecto uniformizador del operador seleccion, y evita la convergencia hacia el primer o ptimo
local encontrado por el algoritmo. Por tanto, la mutacion asume tareas diversificadoras de

la poblacion, y permite trasladar la busqueda

hacia zonas alejadas del espacio de soluciones
(figura 2.16).

y reemplazo. En la naturaleza los individuos mejor adaptaEl proceso de seleccion

probabilidades de reproducirse y de sobrevivir durante varias generados al medio tienen mas
ciones. De esta manera mantienen y diseminan en la poblacion las estructuras geneticas que
dificultades para
desarrollan fenotipos exitosos. Los individuos menos aptos encuentran mas
reproducirse y su esperanza de vida es menor. A este proceso se le denomina Seleccion Natural
y permite que cada generacion mejore su adaptabilidad con respecto a las generaciones precedentes, dando lugar a la evolucion de las especies. En la analoga con la Naturaleza planteada
en la estructura y principios de funcionamiento de los algoritmos geneticos, dicho proceso de
Seleccion Natural se realiza mediante dos procedimientos: la seleccion propiamente dicha y el
reemplazo.

El proceso de seleccion elige un subconjunto de la poblacion formado por aquellos indi por diversas vas, la siguiente generacion. Los procesos en los que
viduos que conformaran,
interviene la seleccion son: la generacion de la poblacion de criadores, la seleccion de los supervivientes y, en caso de que se emplee una estrategia elitista, la seleccion de la e lite.

Aunque existen multiples

variantes, en el Algoritmo Genetico Basico,
el proceso de se de criadores. De
leccion comienza seleccionando a los individuos de la poblacion que serviran
entre estos individuos (que forman una poblacion auxiliar o intermedia) se selecciona a los
Este conjunto de operaciones es la Seleccion propiamente
que efectivamente se reproduciran.
dicha. El proceso de escoger aquellos individuos de entre la poblacion intermedia y los descen la nueva generacion por sustitucion de la antigua, se denomina Reemdientes que formaran

de literatura
2. Revision

plazo. La seleccion puede llevarse a cabo entre otros, por los siguientes tipos de muestreo:

directo, aleatorio simple, estocastico,

diferenciado, conservador, excluyente, por valor esperado, por contadores y universal o por ruleta.

La seleccion adecuada de los parametros

descritos para los AE no es una tarea trivial.
de
Existen interacciones evidentes entre la actuacion de los operadores geneticos y el tamano
la poblacion, que convierten a la parametrizacion en un complejo problema de optimizacion no

fuertemente del tipo de funcion

lineal [134136]. Los parametros
adecuados dependen ademas
de aptitud empleada y, por tanto, del problema a resolver. En el actual estado del arte de la

teora relacionada con los AG, la seleccion de los parametros

se realiza empleando la tecnica de ensayo y error, es poco eficiente, consume mucho tiempo y los resultados obtenidos son
poco confiables, por lo que el tema de la adecuada parametrizacion de los algoritmos geneticos
esta lejos de ser resuelto. Lo anterior pone de manifiesto la evidente necesidad de buscar es
trategias metodologicas y sistematicas
para resolver de forma eficiente los problemas planteados.

Las teoras basadas en tecnicas de IA, como las RNA y/o AE, han brindado una alter
nativa a la computacion clasica
para aquellos problemas en los que los metodos tradicionales
exhan entregado resultados no muy convincentes o poco convenientes. Las aplicaciones mas

itosas de las RNAs son: procesamiento de imagenes

y de voz, reconocimiento de patrones,

planeacion, interfaces adaptivas para sistemas hombre/maquina

(robotica), prediccion, con

trol y optimizacion, filtrado de senales.

En los ultimos
anos,
se ha intentado aplicar la teora

de RNA en al area
de investigacion de la ciencias nucleares, en el ambito
que tiene que ver

con resolver el problema de la reconstruccion de espectros de neutrones y el calculo

de dosis
equivalentes por medio de las lecturas tomadas con el Sistema Espectrometrico de Esferas
Bonner (SEEB). [54, 55, 6265, 97104]

de literatura
2. Revision

2.2.

Espectrometra de Neutrones
Los campos de neutrones se encuentran presentes en particular en el interior de las con-

tenciones de las instalaciones nucleares experimentales o de potencia y en las unidades de

pretratamiento de combustible nuclear [6668, 70, 72, 80, 137, 138]. Se encuentran ademas
sentes en la radiacion cosmica y en la proximidad de los aceleradores de partculas, en particular de aquellos utilizados en algunas aplicaciones medicas. Existe una constante necesidad
en determinar la dosis equivalente de neutrones debido a que diferentes fuentes de neutrones

pueden impactar en las condiciones de trabajo. La medicion y calculo

de la fluencia de espectros de neutrones son un factor clave en la dosimetra para la proteccion radiologica de
neutrones [69, 92, 139143].

Los neutrones ya provengan de fuentes naturales o como los producidos por actividades
humanas, cubren un amplio margen de energas que se extienden desde pocas decimas de

eV o menos, hasta varios centenares de MeV [70, 73, 144]. La variacion rapida
de la seccion

eficaz con la energa del neutron, ha incitado a clasificar a los neutrones en categoras segun
su energa cinetica y tipo de interaccion. Si bien los lmites que definen esta clasificacion aun
sujetos a discusion, los neutrones se clasifican de acuerdo a su energa, ya que el tipo
estan
de reaccion que sufre el neutron depende directamente de e sta [82, 84]. El cuadro 2.1 muestra
una clasificacion aceptada de los neutrones atendiendo a su energa.
Cuadro 2.1: Clasificacion de los neutrones
Neutrones

Rangos de energa

Termicos

< 0.4 eV

Intermedios

0.4 eV - 10 keV

Rapidos

10 keV - 10 MeV

Relativistas

> 10 MeV

Los neutrones termicos son aquellos que se encuentran en equilibrio termico con los

nucleos
del medio donde se encuentran, se les puede aplicar las nociones de la teora cinetica
de los gases y su distribucion sigue la estadstica de MaxwellBoltzman; los neutrones inter

medios resultan de la colision elastica

de los neutrones rapidos
en materiales de bajo numero

de literatura
2. Revision

atomico, la distribucion de e stos neutrones es proporcional al inverso de su velocidad; los neu

trones rapidos
son aquellos que tienen una energa superior a unas decenas de keV, valor utilizado como lmite por debajo del cual los instrumentos para la deteccion de estos neutrones
resultan inadecuados; finalmente, se consideran relativistas todos los neutrones cuya energa
es superior a los 10 MeV.

Como se aprecia en la figura 2.17, la ausencia de carga electrica del neutron le confiere
un elevado poder de penetracion y la posibilidad de interaccionar con la materia de manera
distinta a la de las partculas cargadas ( y ) y radiacion electromagnetica () [145]. Dadas las

dimensiones ocupadas por los nucleos,

e stas interacciones tienen una seccion eficaz pequena
y producen, excepto casos particulares, una escasa perdida de energa, lo que conduce a que

sea necesario un gran numero

de colisiones para que la energa del neutron disminuya de
manera significativa [6668, 70, 72]. Esta circunstancia hace que los neutrones en conjunto
penetrantes que las partculas cargadas y puedan de esta
con la radiacion , sean mucho mas
forma atravesar espesores importantes de materia.

Figura 2.17: Poder de penetracion de partculas radiactivas

Debido a la ausencia de carga electrica del neutron y a la manera en que interacciona

con la materia, no existe un detector de neutrones rapidos

de registro directo (neutrones con
energa cinetica mayor que 200 KeV); por lo que se hace necesario hacerlos pasar por un medio
moderador rico en hidrogeno, en el cual el neutron pierde energa cinetica con el fin de que la
energa del neutron alcance rangos que lo hagan detectable (energa menor que 0.2 MeV). La
principal causa por la cual los neutrones pierden energa en el medio moderador es por coli

siones elasticas
y en menor grado por colisiones inelasticas
[82, 84].

En la figura 2.18, se puede observar que los neutrones rapidos

se frenan dentro del mod-

de literatura
2. Revision

erador y llegan al detector en estado termico, mientras que los neutrones inicialmente termalizados son parcialmente capturados dentro del moderador y no alcanzan el detector. Au
mentando el diametro
de la esfera moderadora, se desplaza el pico de sensibilidad del sistema
haca las energas altas y de esta manera es posible efectuar una espectrometra de neutrones.
Tanto para la dosimetra como para la espectrometra de neutrones, se utilizan estas interacciones, ya que al producir partculas cargadas, e stas inducen en la instrumentacion por medio

de la ionizacion, senales
electricas a partir de las cuales y mediante procedimientos adecuados,
se remonta a la energa del neutron detectado.

Figura 2.18: Deteccion de neutrones empleando medios moderadores

Los neutrones constituyen una radiacion indirectamente ionizante que provoca a igual
importantes que los producidos por los fotones. En la
dosis absorbida, efectos biologicos mas
actualidad los instrumentos de radioproteccion operacionales tienen respuestas muy dependientes de la energa y del espectro neutronico, y sobre todo de la presencia en el mismo, de un
componente importante de neutrones de energas intermedias. Todo ello hace difcil la interpretacion de las indicaciones de los instrumentos utilizados para cuantificar la energa de los
neutrones, si no se conoce la distribucion espectral de fluencia (espectro) de los neutrones en
los puntos de medida.

2.2.1.

Espectrometra de Neutrones con el SEEB

Como se menciono en la seccion 2.2, la medicion y en particular la determinacion de espectros de neutrones no es una tarea trivial, debido principalmente a la ausencia de carga
electrica que e stos presentan [146148]. Las tecnicas utilizadas para la deteccion de los neutrones, se describen ampliamente en numerosas referencias [6670,72,73,7588,92,149159].
Las dificultades especficas de la espectrometra de neutrones provienen de la complejidad y
variedad de las interacciones de los neutrones con el tejido humano, interacciones que presen-

de literatura
2. Revision

tan secciones eficaces muy diferentes en funcion de la energa. El resultado es que para un
se trate de neutrones termicos o rapidos,

mismo valor de fluencia de neutrones, segun

la dosis
absorbida puede llegar a ser un orden de magnitud diferente, diferencia que puede ser aun
mayor para la dosis equivalente. Esta dependencia tan fuerte con la energa de los neutrones
no ha podido ser replicada de modo plenamente satisfactorio por los metodos de medida hoy
da existentes y es la principal causa de incertidumbre en su respuesta, en mayor proporcion
que la causada por la sensibilidad tambien limitada de esos metodos.

Aunque existe un amplio rango de dispositivos utilizados en la espectrometra de neu

de categoras gentrones [146148], la mayora se pueden agrupar en un numero
pequeno
sobresaliente, destacandose

erales, cada una de ellas basada en una tecnica comun

entre e stos
el sistema Espectrometrico de Esferas Bonner (SEEB) [8991, 9193, 93, 160174], el cual, como se aprecia en la figura 2.19 (a y b), esta compuesto de un detector de neutrones termicos

y un conjunto de esferas de polietileno de alta densidad, cuyos diametros

varan de 1 a 18
pulgadas.

(a) Detector de neutrones

(b) Esferas Bonner

Figura 2.19: Sistema Espectrometrico de Esferas Bonner

La tecnica de las esferas Bonner, involucra recolectar pulsos electronicos que surgen de

las interacciones entre los neutrones y los atomos

de 6 Li en un detector de Ioduro de Litio

de literatura
2. Revision

dopado con Europio, 6 LiI (Eu), con el detector en diferentes configuraciones de conteo en el
campo de irradiacion de neutrones. Como se observa en la figura 2.20, la combinacion detectormoderador tendra una respuesta diferente a los neutrones en funcion de la energa.

A partir de la tasas de conteo tomadas con las esferas, es posible reconstruir el espec proporciona informacion acerca de la distribucion de energa de los
tro de neutrones, el cual
neutrones incidentes. Como puede observarse en la figura 2.20, en el SEEB cada detector se
del modercaracteriza por una funcion de respuesta. A medida que se incrementa el tamano
ador, el pico de la funcion de respuesta cambia hacia neutrones de energa mayor.

Figura 2.20: Combinacion detector-moderador

Cuando se utiliza el SEEB, se miden las tasas de conteo que los neutrones producen en
de la esfera ocasiona que el siscada uno de los detectores del sistema [175, 176]. El tamano
tema tenga una eficiencia de deteccion de neutrones en funcion de la energa, as, la matriz de
ntimamente relarespuesta del sistema, las tasas de conteo y el espectro de neutrones estan
cionados a traves de la ecuacion integro-diferencial de Fredholm de primer tipo [177]. Cuando

neutrones de energa Ei alcanzan la esfera de diametro

D j los neutrones pierden energa y
que se traduce en una tasa de conteo C j [94]. El
alcanzan al detector produciendo una senal

incrementar el diametro
de las esferas permite que se puedan detectar neutrones de energas
mayores a las termicas. Al conjunto de eficiencias en funcion de la energa de los neutrones se

le llama matriz de respuesta, la relacion entre las tasas de conteo, en funcion del diametro
de
la esfera C (D), la matriz de respuesta R(D, E ) y el espectro de las muestras E (E ), se muestran
en la ecuacion 2.7 [175, 176].

de literatura
2. Revision

C (D) =

Z Emax
Emin

R(D, E )E (E )dE

(2.7)

Esta es la ecuacion integro-diferencial de Fredholm de primer tipo, que en su version

discreta se indica en la ecuacion 2.8.

Cj =

Ri, j i

(2.8)

i=1

Donde C j es la tasa de conteo del j-esimo detector, Ri, j es la matriz de respuesta del je simo detector de neutrones en el i-esimo intervalo de energa, i es la fluencia de neutrones

en el i-esimo intervalo de energa y m es el numero

de esferas del sistema espectrometrico.

Debido a que el numero

de detectores (esferas) es menor que el numero
de grupos de
energa usados para describir el espectro, la ecuacion 2.8 se convierte en un sistema de ecua

ciones mal condicionado, en estos casos no existe una solucion unica,

se tiene un numero
infinito de soluciones; el seleccionar aquella que tenga significado para el tipo de problema es
procedimiento para la
parte del proceso de reconstruccion, por lo que se debe aplicar algun
deconvolucion del espectro [94, 95, 171, 175, 178200].

La figura 2.21 muestra que al utilizar las cuentas integrales ledas a partir del SEEB es
posible reconstruir el espectro de neutrones empleando varios metodos, entre los que se pueden
mencionar: los metodos Monte Carlo [198, 201205], de Parametrizacion [74, 95, 191, 206] e Iterativos [160, 189, 192, 207209].

De la figura 2.21, se puede observar que las principales dificultades de e stos metodos,
es la alta dependencia de la experiencia del experimentador y de la calidad de la informacion a priori (espectro inicial), que debe ser proporcionada a los mismos, asimismo, habiendo
obtenido el espectro, las dosis se calculan por medio de los coeficientes de fluencia de dosis, que
en la mayora de los codigos de reconstruccion, son coeficientes no actualizados lo que conlleva
a subestimarla, o bien, a obtener una dosis que ya no se utiliza.

de literatura
2. Revision

Figura 2.21: Tecnicas clasicas

para espectrometra de neutrones con el SEEB
Los metodos existentes hoy da empleados en la espectrometra y dosimetra de neutrones presentan limitaciones que pueden ser importantes en sus caractersticas operativas.

Como consecuencia de estas limitaciones, la dosimetra de neutrones, tanto la de area

como la
personal, ha de hacerse con unos niveles de incertidumbre comparativamente mayores que la
a pesar del muy considerable y
dosimetra de fotones o de electrones. Esta situacion continua
continuo esfuerzo investigador llevado a cabo para tratar de encontrar metodos alternativos o
de mejorar los ya existentes.

La espectrometra de neutrones ha tenido que hacer frente durante su desarrollo en es

tas ultimas
decadas a numerosas dificultades, mismas que provienen tanto del dominio de
la teora como de la instrumentacion y en su inmensa mayora se deben a las propiedades
especficas de la interaccion de los neutrones con la materia, lo que explica que la precision
conseguida hoy da dentro del campo de la dosimetra de neutrones, sea netamente inferior al
encontrado en el caso de la radiacion gama.

No obstante las dificultades mencionadas o precisamente a causa de ellas, grupos de in-

de literatura
2. Revision

vestigadores han producido algunos avances significativos en los ultimos

anos.
Recientemente

se han utilizado metodos como los basados en la Maxima

Entropa [210, 211] o en tecnicas de
IA como AG [212214] y/o RNA [54, 55, 6265, 97104, 215217], como metodos alternativos,
mismos que pretenden resolver algunos de los problemas antes mencionados.

Como se observa en la figura 2.22, la reconstruccion de espectros de neutrones a partir

del SEEB por medio de la tecnologa de IA, es una tecnica alternativa que esta ganando popularidad entre los investigadores, ya que ofrece mejores resultados con relacion a los metodos
de solucion tradicionales, empleados para tal fin.

Figura 2.22: Espectrometra de neutrones con el SEEB y tecnicas de IA

recibiendo gran atencion por parte de los inEn la actualidad, las tecnicas de IA estan
vestigadores, debido a los resultados exitosos obtenidos, sin embargo, como puede observarse

en la figura 2.22, en el actual estado del arte que tiene que ver con el uso de la IA en el ambito
existen serios problemas con lo que se refiere a la sede la espectrometra de neutrones, aun

leccion o ptima de parametros

de RNA y AG.

Actualmente los investigadores determinan e stos parametros

por medio de la tecnica de

de literatura
2. Revision

ensayo y error. Lo anterior implica que no es facil

para el disenador
elegir una topologa de
algoritmo o red satisfactorios, sin importar la experiencia que tenga.

Asimismo, debido a que es una tecnologa emergente, la carencia de herramientas cientfi

cas y tecnologicas que estimulen su aplicacion practica
en situaciones reales, como en el caso

del presente proyecto en la espectrometra neutronica, representa un serio obstaculo

para su
desarrollo.

sencilla y eficiente de superar e stas desvenPor consiguiente, debe existir una forma mas
tajas. Un investigador, con o sin experiencia en RNA y/o AG, no debera emplear mucho tiempo

en la seleccion de los parametros

o ptimos, sin embargo, en la practica,
el mayor porcentaje de
tiempo se dedica a e sta actividad, porque se carece de metodos y estrategias en este dominio
del conocimiento.

de las RNA y/o AG no es un problema trivial, y la determiEl procedimiento de diseno

nacion de un conjunto apropiado de valores de los parametros

estructurales y de aprendizaje
de las mismas sigue siendo una tarea difcil.

Las dificultades intrnsecas asociadas con las RNA y los AG hacen difcil aplicar e stas

que permitan aplicar e sta tecnologa de forma agil

y eficiente, en ambientes laborales, experimentales y de laboratorio.

Captulo 3

Materiales y metodos
3.1.

Introduccion
El objetivo primario del presente trabajo de investigacion se centra en dos aspectos fun-

y aplicar tecnicas de optimizacion, metodologidamentales: en primer lugar, el estudiar, disenar

de arquitecturas robustas de RNA capaces de resolver con alto

cas y sistematicas,
en el diseno
rendimiento los problemas relacionados con la espectrometra y dosimetra de neutrones. El
de herramientas tecnologicas que permitan aplicar el conocimiensegundo objetivo es el diseno

to adquirido por las redes disenadas

en ambientes laborales, experimentales, y de laboratorio.

En e sta seccion, se exponen los materiales y metodos empleados en la busqueda

de conseguir
los objetivos planteados.

Como se observa en la figura 3.1, la espectrometra y dosimetra de neutrones empleando la tecnologa de RNA a partir de las tasas de conteo del SEEB realizada en este trabajo,
se divide en varias etapas: en primer lugar, se realiza una etapa de pre-procesamiento de los
datos que conformaran las matrices de entrenamiento de la RNA. En esta etapa del proyecto,

se disenaron
los codigos de apoyo denominados; Codigo que automatiza la creacion de matrices de entrenamiento de RNA empleadas en espectrometra Neutronica, herramienta que
automatiza el proceso de creacion de las matrices de entrenamiento de las redes neuronales

a ser disenadas,
y el Codigo Visual para la ejecucion del codigo MCNP (VMCNPex, de sus
siglas en ingles) [218], que permite realizar simulaciones de forma visual, amigable e intuitiva

del codigo MCNP, en el proceso de rebobinado y calculo

de 13 cantidades dosimetricas en la
espectrometra de neutrones.

3. Materiales y metodos

Figura 3.1: Esquema general de RNA aplicadas en la espectrometra neutronica

Una vez que se construyeron las matrices de datos de entrada y salida de las redes a

ser disenadas,
como se aprecia en la figura 3.2, la siguiente etapa consistio en determinar las
topologas o ptimas de las redes, para lo cual se exploraron dos metodologas: la Metodologa
de Diseno
Robusto de Redes Neuronales Artificiales (MDRRNA) (RDANN, de sus siglas en
de Experimentos (DoE)
ingles), que es una tecnica estadstica basada en la teora del Diseno
[219, 220], y la tecnica basada en Redes Neuronales Evolucionadas por medio de Algoritmos
Evolutivos (RNAE) (EANN, de sus siglas en ingles), una tecnologa hbrida de redes neu
ronales disenadas
por medio de AE [221226].

Figura 3.2: Tecnicas de Optimizacion de RNA: DRRNA y RNAE

3. Materiales y metodos

Con las redes entrenadas mediante las metodologas mencionadas, como puede observarse en la figura 3.3, la siguiente etapa centro su atencion en el desarrollo de herramientas computacionales, mismas que permitieron emplear de forma eficiente, el conocimiento
adquirido por las RNA en situaciones experimentales reales. Paro lo anterior, luego de ex
traer el conocimiento de las redes creadas, se disenaron
los codigos de reconstruccion de
espectros de neutrones basados en la tecnologa de IA conocidos como: Espectrometra y
Dosimetra de Neutrones por medio de Redes Neuronales Artificiales (NSDann, de sus siglas en ingles) [220,227230] y Espectrometra y Dosimetra de Neutrones por medio de Redes
NeuroEvolutivas (NSDEann, de sus siglas en ingles) [231233]. En esta etapa, tambien se
o la herramienta de apoyo denominada Caja de Herramientas para la Espectrometra
disen
y Dosimetra de Neutrones (NSDTB, de sus siglas en ingles) [234236], la cual permite comparar los espectros obtenidos con los diversos codigos de reconstruccion.

Figura 3.3: Codigos para espectrometra de neutrones basadas en RNA

Los codigos mencionados, como se muestra en la figura 3.4, se disenaron

en en una etapa
incial en base a un SEEB con un detector de Ioduro de Litio activado por Europio 6 LiI (Eu),
compuesto de 7 esferas de polietileno para espectros de neutrones expresados en 31 grupos de
energa [227, 228]. Sin embargo, debido a los resultados alentadores obtenidos, una segunda
consistio en modificar los codigos previamente disenados

etapa en este proceso de diseno

en
base a un SEEB con un detector de 3 He, compuesto de 12 esferas de polietileno para espectros

3. Materiales y metodos

de neutrones expresados en 72 grupos de energa [220, 224, 237].

Figura 3.4: Codigos para espectrometra de neutrones basadas en RNA para un detector de
3 He

Construidos los prototipos de las herramientas computacionales, se realizo una etapa

de optimizacion de los mismos, cuya finalidad consistio en calcular la incertidumbre asociada
con las tasas de conteo obtenidas al emplear el SEEB, utilizando la estadstica de Poission,
e implementada en una rutina de computo [238]. La etapa siguiente consistio en el uso de

las herramientas de computo disenadas,

en ambientes experimentales reales, y comprobar su
eficacia y rendimiento, por ello, en e sta etapa se realizaron experimentos en la reconstruccion
de espectros de neutrones de diversas fuentes de neutrones en distintas condiciones, as como
en aceleradores de partculas de uso medico. Para lo anterior, se utilizaron los datos experimentales obtenidos con los SEEB de la Unidad Academica de Estudios Nucleares (UAEN),
perteneciente a la Universidad Autonoma de Zacatecas (UAZ), Mexico, el Departamento de
y el
Ingeniera Nuclear (DIN), de la Universidad Politecnica de Madrid (UPM) en Espana
Laboratorio de Patrones Neutronicos (LPN), perteneciente al Laboratorio de Metrologa de
Radiaciones Ionizantes (LMRI) del Centro de Investigaciones Energeticas Medio Ambientales

y Tecnologicas (CIEMAT), con sede en Madrid, Espana.

Una vez realizadas las mediciones explicadas en el parrafo

anterior, y como se ilustra en

3. Materiales y metodos

la figura 3.5, en la siguiente etapa se compararon los resultados obtenidos con las herramien

tas disenadas,
contra los resultados obtenidos utilizando las metodologas clasicas
empleadas
en la actualidad en los distintos laboratorios como son el codigo BUNKIUT en el caso de la

UAZ y la UPM y el codigo UMG en el caso del CIEMAT, realizando un analisis

de semejanzas

y diferencias y resaltando las ventajas y desventajas que los sistemas disenados

presentan con

relacion a las metodologas clasicas.

Figura 3.5: Comparacion de metodos clasicos

y los basados en IA en espectrometra neutronica
A continuacion se explica con mayor detalle cada una de las etapas realizadas durante
la presente investigacion.

3.2.

Pre-procesamiento de la informacion
El procedimiento utilizado para obtener los datos de entrenamiento de la RNA fue prop-

uesto por Vega-Carrillo et al [97104]. En forma esquematica

el procedimiento se muestra
en la figura 3.6, en donde se puede observar que es necesario realizar varias etapas antes de

poder iniciar con el proceso de entrenamiento de las redes a ser disenadas,

siendo una de las
mas importantes la etapa de pre-procesamiento de los datos de entrenamiento de la red que

se desea disenar.

3. Materiales y metodos

Figura 3.6: Metodologa utilizada para la obtencion de los datos del entrenamiento de la RNA.
El contar con un conjunto de datos, esto es, las matrices de entrenamiento tanto de entrada como de salida deseada, no es una tarea trivial, ya que como puede apreciarse en la
figura 3.6, este proceso requiere el rebobinado de un conjunto de espectros de neutrones, mediante simulaciones realizadas con el codigo MCNP, tomados de una compilacion realizada por la
Agencia Internacional de Energa Atomica (IAEA, de sus siglas en ingles) [85,86,239]. Asimis
mo, mediante simulaciones MCNP, se requiere realizar el calculo
de 13 cantidades dosimetricas que incluyen de forma general la dosis equivalente ambiental, la dosis equivalente personal
y la dosis efectiva, utilizando los coeficientes de conversion de fluencia a dosis publicados por
el ICRP74 [240]. Lo anterior, involucra invertir una gran cantidad de tiempo, debido principalmente a que cada espectro de neutrones debe ser tratado de forma independiente desde el
principio hasta el fin, en la forma que se describe a continuacion.

Se tomaron todos los espectros contenidos en el compendio de espectros de neutrones

compilado por la IAEA , dentro de los cuales se encuentran espectros producidos por fuentes
isotopicas de neutrones, reactores nucleares, aceleradores de uso medico y de investigacion en
fsica, rayos cosmicos, etc. Estos espectros se encuentran expresados en unidades de letargia
definidos en 60 grupos de energa, por lo que de letargia se convirtieron a espectros
y estan
expresados por unidad de energa, utilizando la siguiente ecuacion 3.1

3. Materiales y metodos

E (E ) = E (u) ln(

Esup
)
Ein f

(3.1)

Donde E (u) es el espectro de neutrones expresado en unidades de letargia, Esup es el

intervalo de energa superior y Ein f el intervalo de energa inferior.

Por cada uno de los espectros expresados en unidades de energa, se crea un archivo de
entrada al codigo MCNP con el objetivo de realizar un proceso de rebobinado por cada uno
de los espectros tomados en consideracion. Lo anterior, tiene como proposito modificar la estructura de energa de los espectros de neutrones y calcular 13 cantidades dosimetricas, de

acuerdo con las caractersticas de la energa de la matriz de respuesta del SEEB para el cual
la RN.
se desea disenar

Si bien la tecnica propuesta por Vega-Carrillo et al ha mostrado ser prometedora, el pro

ceso de rebobinado de los espectros de neutrones y el calculo
de las 13 cantidades dosimetricas,
sufre un gran inconveniente: no se cuenta con herramientas tecnologicas, que hagan eficientes

y agiles
las tareas relacionadas con el procesamiento de los datos. Actualmente las tareas
involucradas en e sta etapa se realizan de forma completamente artesanal, utilizando proce
sadores de texto y hojas de calculo,
en donde el investigador copia y pega la informacion de
una aplicacion a otra, ya sea con el teclado o con el raton, lo que implica invertir grandes
cantidades de tiempo, lo que implica el riesgo de cometer errores de tipo humano por fatiga o
falta de concentracion. Lo anterior puso en evidencia la necesidad de contar con herramientas
computacionales que agilizaran y facilitaran la realizacion de las tareas descritas.

En base a lo expuesto anteriormente y debido a la gran cantidad de informacion por

manipular, al numero
elevado de simulaciones MCNP por realizar, ya que se tiene que crear
un archivo de entrada por cada uno de los espectros que se utilizan en la matriz de datos de
entrenamiento de la RNA y, a que la ejecucion del codigo se realiza bajo entornos de consola
de comandos en Unix o MS-DOS en Windows, como se aprecia en la figura 3.7, el proceso de
rebobinado de espectros resulta ser una tarea ardua y difcil, que demanda grandes cantidades
de tiempo. Un proceso de rebobinado tpico de los espectros compilados por la IAEA sin el
auxilio de herramientas tecnologicas, podra abarcar un periodo de tiempo de 7 a 9 meses .

3. Materiales y metodos

Figura 3.7: Ejecucion del codigo MCNP en el entorno MS-DOS de Windows

Al observar la figura 3.7, se puede apreciar que el entorno de consola de comandos o MSDOS no es amigable al usuario, ya que se tiene que escribir un comando por cada accion que se
desea realizar, proceso que se vuelve en extremo lento y tedioso, como en el caso del presente
proyecto, ya que son muchas las acciones a ejecutar sobre una gran cantidad de archivos que
se manipulan, perdiendose muy pronto el control sobre ellos.

El proceso de rebobinado tradicional de espectros de neutrones y calculo

de dosis equivalente por medio del codigo MCNP, involucra tres etapas fundamentales. En primer lugar, se
requiere crear un archivo de entrada al codigo MCNP por cada uno de los espectros utilizados
de la compilacion realizada por la IAEA. Cada archivo de entrada debe contener las carac se desea disenar
la
tersticas de la energa de la matriz de respuesta del SEEB para el cual
RN y 13 cantidades dosimetricas. Este proceso puede tardar de 3 a 4 meses, de acuerdo con
la velocidad con la que trabaje el investigador, ya que al no tener herramientas de apoyo, el
proceso se realiza muy lento.

Debido a que por ser una tecnologa emergente no existen herramientas que permitan
eficientar y agilizar el proceso de creacion de las matrices de entrenamiento de RNA empleadas en la espectrometra neutronica y, dada la necesidad de contar con una herramienta
una interfase grafica

tecnologica que resuelva los problemas descritos, se diseno

de usuario en
el entorno de programacion de LabVIEW denominada Codigo que automatiza la creacion de

3. Materiales y metodos

matrices de entrenamiento de RNA empleadas en espectrometra Neutronica, mostrada en la

figura 3.8.

Figura 3.8: Codigo que automatiza la creacion de matrices de entrenamiento de RNA en espectrometra Neutronica

Esta herramienta permite realizar de forma automatica,

en donde el humano interviene

solo para establecer algunos parametros,

la generacion de archivos de entrada para la simulacion del codigo MCNP con cada uno de los espectros contenidos en la compilacion de la
IAEA. Lo anterior reduce significativamente el tiempo empleado en esta actividad de meses a
segundos.

La segunda etapa en el proceso de rebobinado involucra la simulacion de cada uno de los

archivos de entrada al codigo MCNP creados en la etapa anterior. Por cada archivo simulado,
el tiempo invertido puede ser de algunas horas a varios das por espectro, debido entre otras
razones a que el codigo MCNP se ejecuta sobre una plataforma MS-DOS, lo que hace muy
complicada la manipulacion de archivos. Dadas las multiples desventajas que representa el
o el codigo denomiuso de este entorno en el proceso de rebobinado de los espectros, se disen

3. Materiales y metodos

nado Codigo Visual para la ejecucion del codigo MCNP (VMCNPex, de sus siglas en ingles),
de utilizar creada
mostrado en la figura 3.9. Esta es una herramienta grrafica, intuitiva y facil
para ser ejecutada en ambientes de trabajo Windows, ofrece grandes ventajas en las opera
ciones de manipulacion de la informacion si se compara con la tecnica clasica
previamente
descrita. El uso de esta herramienta, resuelve muchos de los problemas asociados con el metodo tradicional y ahorra cantidades significativas de tiempo, permitiendo reducirlo de meses a
semanas u horas de acuerdo con la plataforma de computo empleada en la simulacion.

Figura 3.9: Entorno grafico

de VMCNPex

El entorno grafico
de VMCNPex permite entre otras cosas, ejecutar las simulaciones MCNP sin la necesidad de tener que escribir comandos MS-DOS. Para ejecutar el programa, el
usuario solo tiene que seleccionar la ruta donde se encuentra almacenado el ejecutable MCNP
se dispone de una herramienta de seleccion, mostrada en la parte
(mcnp.exe), para lo cual
superior derecha de la pantalla. Una vez seleccionada la ruta del ejecutable, el usuario puede
observar en la ventana que aparece bajo la herramienta de seleccion de carpetas, que archivos
se encuentran dentro de la carpeta seleccionada.

La tercera etapa en el proceso de rebobinado implica recuperar la informacion de espectros y dosis rebobinados mediante la simulacion MCNP. Debido a la gran cantidad de archivos
generados en el proceso de simulacion, el proceso de recuperacion de la informacion correspon-

3. Materiales y metodos

diente a los espectros rebobinados y dosis calculadas sin el auxilio de herramientas tecnologicas implica buscar la informacion correspondiente en cada archivo de salida producido por la
simulacion del codigo MCNP. Esta actividad podra abarcar un periodo de tiempo de 3 a 4
meses. Por la razon anterior, se desarrollo en el codigo descrito en la etapa uno, una rutina de

computo que permite extraer en un ambiente grafico

y de forma agil
y eficiente, la informacion
relevante de e stos archivos. lo anterior permite reducir el tiempo empleado en esta actividad
de meses a minutos.

Con la informacion correspondiente a los espectros rebobinados y las 13 dosis calculadas

de cada uno de los espectros de neutrones utilizados a partir de la compilacion relaizada por
la IAEA y simulados con el codigo MCNP, se crearon las matrices de datos de entrada y salida
entrenamiento y prueba de la RNA. La matrz de saliutilizadas en los procesos de diseno,
da esta conformada por los espectros de neutrones rebobinados, expresados en unidades de
energa, y 13 cantidades dosimetricas. Al multiplicar cada espectro de la matriz de datos de
se disena
la correspondiente RNA, se
salida con la matriz de respuesta del SEEB para el cual
calculan las tasas de conteo que este sistema producira. Estos datos se utilizan como matrz
de entrada de la RN.

La tecnica propuesta por Vega-carrillo et al y las referencias analizadas que tienen que
ver con el uso de la tecnologa de RNA en la espectrometra neutronica, ponen en evidencia un
de la arquiserio inconveniente: emplean la tecnica de ensayo y error en el proceso de diseno
de
tectura de las redes neuronales. Por la experiencia obtenida en etapas tempranas de diseno
arquitecturas de RNA, se observo que utilizar la tecnica de ensayo y error implicaba invertir

de las redes,
grandes cantidades de tiempo en la busqueda
de los parametros
o ptimos de diseno
esto sin considerar el tiempo empleado
que abarcaban periodos de tiempo de mas de un ano,
en el proceso de rebobinado para la creacion de las matrices de entrenamiento explicado pre
viamente. La tecnica de ensayo y error no proporciona certeza alguna de encontrar en algun
momento particular la mejor topologa de red, debido a la falta de herramientas cientficas,

que pudieran ser aplicadas en este proceso.

metodologicas y sistematicas

Con la metodologa antes mencionada, el usuario o experimentador debe reconstruir los

3. Materiales y metodos

espectros de neutrones utilizando el entorno de Matlab, lo que hace muy difcil su implementacion en ambientes experimentales reales. Lo anterior se debe a que se requiere tener
instalado el programa Matlab y, posteriormente se requiere tener el archivo fuente que con

tenga la arquitectura de red disenada

y los pesos sinapticos
de misma.

de lo antes mencionado, es indispensable que para operar e sta herramienta, el

Ademas
usuario cuente con bases solidas en lo que se refiere a la teora de RNA as como en la operacion general del entorno de Matlab. Para aplicar con e xito e sta tecnologa en el proceso de
tener
reconstruccion de espectros de neutrones por medio de e sta tecnica, se requiere ademas
conocimiento en el uso de la herramienta especializada que se utiliza en Matlab para trabajar
con RNA (toolbox de RNA). Lo anterior, para un usuario no experimentado implica un serio
inconveniente.

Otro inconveniente con el uso de e sta metodologa, es el hecho de que para realizar el

calculo
de las dosis equivalentes, se requiere utilizar una RNA diferente. Lo anterior implica
aplicar la tecnica de ensayo y error para determinar la topologa de red, tecnica que presenta
grandes inconvenientes. Para reconstruir los espectros y calcular las dosis es indispensable
que el usuario cuente con los dos archivos fuente que contengan sendas redes entrenadas para
cada caso particular, y ejecutar en procesos separados cada una para obtener los resultados
buscados.

de e sta
Un inconveniente adicional se presenta cuando se desea comparar el desempeno

tecnologa contra el obtenido con metodos clasicos.

Por lo anterior, la siguiente tarea consistio en investigar, analizar y aplicar estrategias

o ptimo de RNA, mismas que se detallan a continuacion.
de diseno

3.3.

de RNA
Tecnicas
de optimizacion
Aunque los trabajos propuestos por Vega-Carrillo et al establecen las bases metodologi-

cas para realizar el pre-procesamiento de la informacion, carecen de una metodologa sis

de las arquitecturas. Al igual que otros autores quienes han propuesto
tematica
para el diseno

3. Materiales y metodos

el uso de las RNA en la espectrometra y dosimetra de neutrones, la determinacion de los

parametros
de RN se realiza por medio de la tecnica de ensayo y error. En este proceso de
convencional de RNA, una practica

es comenzar con una arquitectura, utilizandiseno

comun

do el juicio de alguien que haya disenado

una red semejante al problema que se desea resolver
o haciendo uso de algunas pautas disponibles; posteriormente y en base a la suposicion de
las pautas anteriores se supone una arquitectura mejorada, comparando la arquitectura ac
es comenzar con una arquitectura arbitraria y camtual con la previa. Otra practica
comun

manteniendo constantes los otros

biar sucesivamente el nivel de cada parametro
de diseno,

parametros
en sus niveles de inicio. Posteriormente, el mejor nivel de cada uno se determina
entre sus niveles iniciales y los niveles finales realizados en los sucesivos cambios. Este ciclo
de suposicion y comparacion se repite hasta que se encuentra una arquitectura con una mejo
ra, que en la mayora de las ocasiones es mnima. Claramente se observa que esta practica
es
ineficiente, consume mucho tiempo y en la mayora de los casos falla en la identificacion de
persiste la necesidad de contar con bases cientficas y
una arquitectura aceptable, con lo cual
y optimetodologicas que permitan resolver los problemas descritos previamente en el diseno
mizacion de arquitecturas de RNA.

Aunque no existe una respuesta universal para tales topicos, en el presente trabajo se
exploraron dos metodologas de optimizacion de redes neuronales Artificiales: la MDRRNA y
o ptimo de las
las tecnicas hbridas basadas en la combinacion de RNA y AE para el diseno
arquitecturas, que permitieron reducir significativamente el periodo de tiempo empleado en el
de la arquitectura de red y en el entrenamiento y prueba de las mismas, de meses a mindiseno
utos, lo que permitio probar grandes cantidades de arquitectura diferentes de forma eficiente
en muy poco tiempo. El tiempo excedente, como se explica mas adelante, se aprovecho para
desarrollar herramientas tecnologicas, orientadas al usuario final, que tuvieran la capacidad

de aprovechar el conocimiento adquirido por las RNA disenadas,

en la solucion de problemas

practicos
y reales por medio de una interfase grafica
de usuario, facil,
rapida
e intuitiva en
ambientes laborales, experimentales y de laboratorio.

3. Materiales y metodos

3.3.1.

Robusto de Redes Neuronales Artificiales

Diseno

Con e sta metodologa, propuesta, disenada

y automatizada en un codigo computacional
2005 [219], se pueden disenar
entre otras, RNA multicapa
por Ortiz-Rodrguez et al en el ano
del tipo propagacion hacia delante con un algoritmo de aprendizaje de propagacion inversa con

aprendizaje supervisado, como las disenadas

en el dominio de la espectrometra neutronica.
Por tanto, este tipo particular de redes necesitan un conjunto de datos de entrenamiento que
le describan cada salida ( pQ ) y su valor de salida esperado (tQ ), como se indica en la ecuacion

3.2. El entrenamiento de una RNA de este tipo implica actualizar los pesos sinapticos
de tal

forma que se reduzca el error medio cuadratico

(MSE, de sus sigla en ingles), expresado en
la ecuacion 3.3, entre las salidas de la red tQ y la respuesta actual que este siendo modelada

pQ . Para lo anterior se utiliza el algoritmo de aprendizaje de propagacion inversa, el cual

debe ajustar los parametros

de la red para minimizar el MSE. Lo anterior mide de forma
de la red, esto es, la calidad del aprendizaje alcanzado por la misma.
cuantitativa el desempeno

( p1,t1), ( p2,t2), ..., ( pQ ,tQ )

(3.2)

donde pQ es una entrada a la red y tQ es la correspondiente salida deseada para el patron

q-esimo.

El algoritmo de aprendizaje de propagacion inversa para RNA multicapa, realiza una

labor de actualizacion de pesos y ganancias con base en el MSE que se muestra en la ecuacion
efectua
N comparaciones entre el espectro original (tQ ), y el reconstruido por la
3.3, el cual
RNA (pQ ), hasta que se obtiene un valor mnimo especificado por el usuario [45].
MSE =

1 N
(E (E )RNA
E (E )Original
)2
i
i
N i

(3.3)

donde N es el numero
de experimentos, E (E )Original
es el espectro original y E (E )RNA
es
i
i
el espectro reconstruido con la RNA.

de RNA, ya que su proposito es

La MDRRNA ofrece beneficios potenciales en el diseno
de RNA, de tal forma que
determinar experimentalmente las selecciones o ptimas de diseno
del sistema, MSE, ademas
de ser o ptimo, sea robusto a diferentes variables de
el desempeno

3. Materiales y metodos

ruido. [Ortiz-Rodrguez et al., 2005] Con el uso de un procedimiento sistematico

como el diseno
robusto, se evita utilizar el poco eficiente metodo de ensayo y error.

Como se observa en la figura 3.10, la MDDRNA es un metodo sistematico

y experimental,

compuesto de 4 fases principales: planeacion, experimentacion, analisis

y confirmacion. En
este trabajo, la MDRRNA se utilizo para determinar los valores o ptimos de las variables de
de una RNA capaz de reconstruir espectros de neutrones, as como calcular 13 dosis
diseno

equivalentes, utilizando unicamente

como entradas, las tasas de conteo obtenidas a partir de
y la capacidad de generalizacion de la red, bajo
un SEEB, maximizando con ello el desempeno
diferentes condiciones de ruido. Dicha metodologa se utilizo para determinar la arquitectura
de red de un SEEB basado en un detector de 6 LiI (Eu) y la de un SEEB basado en un detector
de 3 He.

robusto de RNA
Figura 3.10: Metodologa de diseno

En el procedimiento sistematico
aplicando la MDRRNA, se identifican los efectos es de la red neuronal dentro de la meditadsticamente significativos de las variables de diseno

cion de la robustez (relacion senal-ruido).

Determinar las caractersticas de calidad a ser optimizadas es el primer paso de la

MDRRNA, en la etapa de planeacion. Las caractersticas de calidad son un parametro

cuya

variacion tiene un efecto crtico en la calidad de la RNA a ser disenada.

En la MDRRNA, el
(MSE) es la respuesta a ser observada en la red. Entre los diversos parametros

desempeno

3. Materiales y metodos

de la red de propagacion inversa, se seleccionan cuatro variables

que afectan el desempeno
considerando tres niveles para cada una. Para disenar
la estructura de una RNA
de diseno,
multicapa de propagacion inversa con aprendizaje supervisado empleando la MDRRNA, se se

leccionan como parametros

de aprendizaje: el numero
de neuronas en la primera (A) y segunda
capas ocultas (B), el momento (C) y la tasa de aprendizaje (D), como se aprecia en el cuadro
3.1.
Cuadro 3.1: Variables de diseno
y sus niveles

V.Diseno

Nivel1

Nivel2

Nivel3

Donde A son las neuronas en la primera capa oculta, B las neuronas en la segunda capa
oculta, C el momento y D la tasa de aprendizaje.

En el interior y en el exterior de la red existen fuentes de perturbacion o ruido. Entre

de una RNA se pueden menlas principales fuentes de perturbacion que afectan el desempeno

cionar el generado por la incializacion aleatoria pesos sinapticos

y el debido al ruido aportado
por el porcentaje de datos, a partir de las matrices de datos de entrenamiento generados en
la etapa anterior, que deberan contener los conjuntos de entrenamiento y prueba, respectivamente. [Ortiz-Rodrguez 2005]. Por tal motivo, en la MDRRNA se incluyen e stas variables,
seleccionando dos niveles para cada una.

del porcentaje de datos que deberan asignarse para el conjunto de entreEl tamano
del porcentaje de datos que deberan asignarse para el
namiento comparado con el tamano
conjunto de prueba (V ), se selecciono de tal forma que el 60 % y el 80 % del total de datos se
destinaron para el entrenamiento, dejando los datos restantes para los conjuntos de prueba,
como se muestra en el cuadro 3.2. Para crear los datos de la seleccion aleatoria de los conjuntos
de entrenamiento y prueba (W ), el conjunto de datos se dividio aleatoriamente en dos subcon
juntos de acuerdo con los valores de V usandolos
como conjuntos de entrenamiento y prueba.

3. Materiales y metodos

para generar un conjunto diferente de datos de

Este procedimiento se repitio una vez mas
entrenamiento y prueba.
Cuadro 3.2: Variables de ruido y sus niveles

V.Diseno

Nivel1

Nivel2

6:4

8:2

Entr1/Prue1

Entr2/Prue2

de datos de entrenamiento-prueba y W la seleccion

donde V es la relacion del tamano
aleatoria de los conjuntos de entrenamiento-prueba.

experimental empleando la MDRRNA, se utiliza el arreglo ortogoEn la etapa de diseno

nal cruzado L9 (34 ) - L4 (22 ), mostrado en el cuadro 3.3. En cada combinacion de las condiciones
y ruido, se entrenan y prueban las correspondientes redes, observando el desempeno

de diseno
de cada una. Empleando e sta metodologa, cada red se entrena y prueba en 36 ocasiones, real

izando posteriormente un analisis

de la varianza sobre la relacion senal-ruido
del desempeno
de las redes, calculada para cada uno de los experimentos realizados.

La etapa de analisis
implica que en la i-esima ejecucion del arreglo ortogonal disenado,

se calcula la relacion senal-ruido

(R-S/R) con base a los resultados obtenidos en la ejecucion

de los experimentos. Las R-S/R se analizan por medio de la tecnica del analisis
de la varianza
(AOV, de sus sigla en ingles), con el fin de comparar todos los valores entre s, para determinar
Para el diseno

las diferencias significativas y los niveles o ptimos de las variables de diseno.

de la RNA, se toma en cuenta la medicion de la exactitud de la clasificacion de los datos de
prueba con la R-S/R; donde la relacion deseable de S-R es quella que tiene el mayor valor.

En la etapa de confirmacion, el analisis

de la varianza se utiliza para la construccion
en cuestion, verificando la exactitud de este modede un modelo estadstico para el diseno
se calculo y
lo a traves de varios experimentos. La respuesta esperada para el mejor diseno
de las redes se realizaron exconfirmo por medio de las pruebas de confirmacion. En el diseno

perimentos de confirmacion, seleccionando los parametros

con los mejores niveles, entrenando
y probando las mismas.

3. Materiales y metodos

Cuadro 3.3: Arreglo Ortogonal cruzado L9 (34 ) - L4 (22 )

Exp.No.

Prom

R-S/R

de la RNA que se modela,

Si bien la MDRRNA permite optimizar el proceso de diseno

ahorrando una gran cantidad de tiempo y obtenido multiples

redes con alta capacidad de gen
eralizacion, un inconveniente que se presento despues de haber disenado
las correspondientes
arquitectura de red, es que no se contaba en ese momento con una plataforma de computo

que permitiera poner a prueba la calidad de las redes disenadas

en ambientes experimentales
herramientas que satisficieran las necesireales. Lo anterior sugera la necesidad de disenar
dades mencionadas.

Las RNA han sido bien reconocidas por su capacidad de aproximacion, una vez proporcionados datos disponibles de entrada y salida. No obstante, el algoritmo de entrenamiento
convencional utilizado en el desarrollo del modelo de RNA siempre encuentra dificultades
para converger en un minio global. Debido a lo anterior, los AE se introdujeron en el modelado

de RNA orientandolos
hacia la mejora de la exactitud y robustez del modelo de la red.

3. Materiales y metodos

3.3.2.

Redes Evolucionadas por Algoritmos Evolutivos

El entrenamiento de una RNA implica actualizar los pesos sinapticos

de tal forma que
se reduzca el error entre las salidas de la red y la respuesta actual que este siendo modelada.

de RNA, se debe al hecho de que la busqueda

Uno de los obstaculos

principales en el diseno

o ptima de los pesos sinapticos

de la red, es altamente dependiente de los pesos iniciales y si
localizados cerca de un mnimo local, el algoritmo podra quedar atrapado, lo que
e stos estan
de la red que este siendo disenada.

podra ocasionar un pobre desempeno

El proceso de entrenamiento del modelo RNA-AE (RNAE) tiene alguna similitud con el
modelado de RNA por medio de la MDRRNA. En la MDRRNA, los pesos se actualizan de
acuerdo con la funcion del MSE, ecuacion 3.3, para ajustar el par de datos entrada salida.

de la
En el caso de la metodologa RNAE, tambien se toma en cuenta como desempeno
RNA la funcion MSE, sin embargo, la principal diferencia entre e stas dos metodologas de op
timizacion es el tipo de busqueda,
ya que los AE no requieren ninguna informacion de la RNA
que esta siendo modelada como: el numero de capas ocultas, las neuronas por capa, tasa de
aprendizaje, momento etc.

La idea general al combinar RNA y AE se ilustra en la figura 3.11, en la cual puede

observarse que trata de un proceso cclico.

3. Materiales y metodos

de RNAE
Figura 3.11: Metodologa general de diseno
de RNAE se basa en
A partir de la figura 3.11 se puede observar que el metodo de diseno
un proceso cclico fundamental, el cual consiste en:
1. Crear una poblacion inicial de genotipos, (representaciones geneticas de las RNA).
2. Construir las RNA, fenotipos, en base a los genotipos.
(MSE).
3. Entrenar y probar las RNA para determinar su desempeno
4. Comparar la idoneidad de las redes y mantener las mejores.
5. Seleccionar aquellas redes en la poblacion que sean las mejores, descartando aquellas
que no lo sean.
6. Renovar la poblacion de redes con las seleccionadas en el paso anterior
7. Formar parejas con los genotipos de las RNA.
8. Aparear los genotipos a traves del intercambio de genes, caractersticas, de la red.
metodo aleatorio y,
9. Mutar los genotipos por algun

3. Materiales y metodos

criterio de detencion
10. Regresar al paso 2 y continuar con este proceso hasta alcanzar algun
o detenerlo de forma manual.
A traves del proceso descrito, sobreviven las mejores redes y se transmiten sus caractersticas a las futuras generaciones que se combinan con las de otras redes con el objetivo de

encontrar redes cada vez mejores para el problema considerado. Esta

capacidad de busqueda
efectiva que que la busqueda

genetica es mucho mas

aleatoria, ya que el proceso genetico de
recombinar las caractersticas mejora sustancialmente la velocidad al identificar RNA con alto

desempeno.

En general, como se aprecia en la figura 3.12, el AE utiliza cuatro etapas fundamentales

para obtener los valores o ptimos de una RNA:

de RNA por medio de AE

Figura 3.12: Metodologa de diseno

1. Generar una poblacion inicial de pesos aleatorios y construir las correspondientes arquitecturas de RNA

3. Materiales y metodos

2. Evaluar las RNA utilizando la poblacion de pesos creada en el punto anterior. Esto se
realiza a traves de calcular el error de entrenamiento (MSE) y asignarlo al valor de la
funcion de fitness de acuerdo a que tan buenas fueron esas soluciones
3. Seleccionar los padres de la manipulacion genetica y crear una nueva poblacion de individuos
4. Designar como resultado la mejor RNA que aparecio en cualquiera de las generaciones
En vista de que en la actualidad no existen herramientas tecnologicas especficas en el
dominio de la espectrometra de neutrones que permitan aplicar la metodologa de RNAE, en
entrenamiento y prueba de las RNA evolucionadas por medio de AE,
el proceso de diseno,
se utilizo un programa de computo comercial denominado NeuroGenetic Optimizer (NGO),
mostrado en la figura 3.13, desarrollado por BioComp Systems [241]. Lo anterior sugiere la
necesidad de desarrollar una herramienta a la medida para este dominio del conocimiento. En
la actualidad se realizan trabajos tendientes a resolver este inconveniente.

Figura 3.13: Ventana principal de NGO

de RNA involucra tres
El procedimiento general de trabajo al utilizar NGO en el diseno
fases principales.
La primera fase se relaciona con el ingreso de las matrices de datos de entrenamiento de

la RNA. En este punto, el disenador

utiliza NGO para cargar un archivo con extension
*.txt o *.xls, que contenga la informacion mencionada. Posteriormente se indica que parte
a los datos de salida.
de informacion corresponde a los datos de entrada y cual

3. Materiales y metodos

difcil, esta relacionada con la seleccion

La segunda fase, considerada la etapa la mas

de las caractersticas de la RNA a ser disenada,

as como de los parametros
del AG
que se emplea para ello. La mayor dificultad de e sta etapa radica en la falta de bases

metodologicas para la seleccion o ptima de los parametros

del AG. Como se indico en
del AG depende ampliamente de la o ptima seleccion de los
la seccion X, el desempeno

de este. Para ello se emplea la tecnica de ensayo y error, lo que

parametros
de diseno
de RNA por este metodo. Lo
implica resolver los problemas mencionados para el diseno

anterior deja abierta un area

de investigacion que presenta un gran interes.

La tercera etapa comprende un proceso de prueba de la RNA disenada

y optimizada en
la etapa anterior. NGO contiene una herramienta que permite poner a prueba la RNA

disenada
introduciendo datos no incluidos en las matrices de entrenamiento y prueba. EL
principal inconveniente de e sta etapa es que la forma como NGO presenta la informacion
al usuario, no cumple con las reglas establecidas para la espectrometra neutronica, por

lo que su analisis
es complicado, teniendo que hacer uso de otras herramientas.

En el cuadro 3.4, se muestran los parametros

requerido por NGO en el proceso de diseno
de la RNAE que sea capaz de reconstruir los espectros de neutrones y calcular 13 cantidades
dosimetricas.
Cuadro 3.4: Parametros

de diseno
con la metodologa RNAE

Parametros
de la RNA

Parametros
del AG

Tipo de RNA

Numero
de Generaciones

de capas ocultas
No. Max.

de la poblacion
Tamano

No. min-max vueltas entren.

Metodo de seleccion

No. entren. por red

Metodo reemplazo de poblacion

Metodo de mutacion

Metodo de aparemiento

de arquitecLa principal y mas seria desventaja en el uso de AE aplicados en el diseno

turas de RNA, es la seleccion de los parametros

de la RNA a ser disenada
y el AE aplicado
para tal fin. En la actualidad se utiliza la tecnica de ensayo y error, que como se ha comentado en secciones anteriores, es poco eficiente, consume mucho tiempo y no produce resultados

3. Materiales y metodos

confiables. A la fecha no existen bases cientficas y metodologicas que permitan resolver este
gran inconveniente. Sin embargo, a partir de la experiencia obtenida al crear la metodologa
estadstica MDRNA, se observa que combinarla con la teora de AE, parece ser una buena alternativa en el proceso de solucion de este problema. En la actualidad se explora la posibilidad
de fusionar las metodologas MDRRAN-AE, creando una base cientfica y tecnologica hbrida,

en la busqueda
de una solucion satisfactoria a los problemas planteados.

Otro inconveniente que existe con el uso de NGO aplicado en la espectrometra y dosimetra
de neutrones, se debe al hecho de que como es un programa de proposito general, la forma de
presentar la informacion obtenida despues de haber seleccionado la RNA o ptima no cumple
con las reglas establecidas para la espectrometra neutronica, sino que se presenta de acuerdo con el formato establecido por el programa. Lo anterior presenta una seria dificultad, ya

que no es posible realizar un analisis

de los espectros reconstruidos y las dosis calculadas, por
una herramienta tecnologica que permita realizar este
lo que surge la necesidad de disenar

analisis.

Enseguida se comentan algunos detalles de las herramientas tecnologicas que fueron

creadas en base a las necesidades que surgieron en el proceso de analisis

de las metodologas
de RNA comentadas en e sta seccion.
de diseno

3.4.

de herramientas tecnologicas

Diseno
El uso de la tecnologa de RNA en la espectrometra y dosimetra de neutrones presenta

de RNA emretos tecnologicos importantes, ya que por un lado las metodologas de diseno
pleadas actualmente son poco eficientes y por otro lado, como es una tecnologa emergente

que esta siendo recien aplicada en e sta area

del conocimiento, no se cuenta con herramientas apropiadas para aplicarlas de forma efectiva. Lo anterior pone de manifiesto que existe
la necesidad de desarrollar nuevas herramientas que permitan resolver los problemas mencionados.

Como se aprecia en la figura 3.14, se emplearon las metodologas MDRRNA y RNAE en

un intento por encontrar soluciones a los problemas actualmente presentes en los procesos

3. Materiales y metodos

y optimizacion de RNA. Ambas tecnicas se basan en principios distintos, aunque

de diseno
optimizar los parametros

persiguen un objetivo comun:

de arquitectura y de aprendizaje de
RN. Sin embargo, debido a lo novedoso de e stas metodologas, y observando que se carece

de herramientas para realizar el procesamiento, analisis

y presentacion de la informacion
de
calculada por las redes en ambientes experimentales reales, se propuso realizar el diseno
varias herramientas computacionales con el objetivo de solucionar los problemas comentados.

Figura 3.14: Desarrollo de Herramientas de computo en base a MDRRNA y RNAE

En la presente etapa de investigacion, como se aprecia en la figura 3.14, el objetivo se
herramientas tecnologicas a la medida, a partir del conocimiento adquirido
centra en disenar

por las redes neuronales previamente disenadas.

Dicho conocimiento se encuentra almacena
do en los pesos sinapticos
de las diferentes capas que componen a la red.

En su mayora, las herramientas disenadas

cuentan con las siguientes caractersticas:

se crearon bajo el entorno grafico

G de programacion de LabVIEW [242244]; operan de for
ma independiente; cuentan con una forma eficiente de entrada y salida de datos; son faciles

de utilizar; los programas se ejecutan de forma rapida,

obteniendo los resultados en cuestion
de segundos; para la operacion de las herramientas, no es indispensable para el usuario, el
que cuente con conocimientos de espectrometra o disimetra de neutrones as como de redes
neuronales; a diferencia de los codigos utilizados en la actualidad, los cuales estan basados en
algoritmos de reconstruccion iterativos, mismos que requieren para la deconvolucion del espectro un usuario experto, as como informacion a priori (espectro inicial), para la operacion

de las herramientas disenadas

en este trabajo no es necesaria tal informacion, ya que unica-

3. Materiales y metodos

mente requiere como datos de entrada, las tasas de conteo provenientes del SEEB, ademas
de las ventajas ya mencionadas. A continuacion se describen las caractersticas generales de
e stas herramientas.

Otra caracterstica importante de las herramientas creadas, es que fueron disenadas

para que el usuario final las encuentre faciles

de utilizar, sin importar la complejidad que
posible
puedan tener las rutinas de programacion que componen a los codigos, evitando lo mas
que para su uso, el usuario cuente con conocimientos cientficos y/o tecnicos muy especializa
dos en las areas
de espectrometra neutronica y/o redes neuronales artificiales, en las cuales
basadas, como sucede con las herramientas clasicas

estan
actualmente utilizadas en los laboratorios de este tipo y que representa un serio inconveniente.

Las herramientas se disenaron

de tal forma que para comodidad del usuario, los codigos

comparten caractersticas muy similares en lo que se refiere al entorno grafico

y en la forma de
trabajar con los programas. Lo anterior propicia que al aprender a utilizar una herramienta,
se esta en condiciones de utilizar las otras en una cantidad mnima de tiempo, debido a las
que comparten. Esto, a diferencia de las herramientas clasicas

similitudes de diseno
basadas
en entornos de consola MS-DOS, permite ahorrar tiempo en el proceso de aprendizaje y permite de igual forma aplicarlas de manera inmediata en ambientes experimentales reales.

3.4.1.

automatica

VMCNPex y generacion
de matrices de entrenamiento de
RNA

En la seccion 3.2, se hizo una descripcion de las herramientas denominadas Codigo que
automatiza la creacion de matrices de entrenamiento de RNA empleadas en espectrometra
Neutronicay Codigo Visual para la ejecucion del codigo MCNP (VMCNPex). El proposito
fundamental de e stas herramientas, fue el de agilizar e incrementar la eficiencia del procesamiento de la informacion utilizada para la creacion de las matrices de entrenamiento de las

redes a ser disenadas

mediante simulaciones con el codigo MCNP. Empleando las herramientas descritas, fue posible reducir el tiempo de creacion de las matrices de entrenamiento de
meses a horas, agilizando en gran medida el proceso general de RNA aplicadas en espectrometra neutronica.

3. Materiales y metodos

3.4.2.

NSDann

La herramienta denominada NSDann (de sus siglas en ingles, Neutron Spectrometry

o a partir del conocimiento
and Dosimetry by means of Artificial Neural Netoworks) se disen

adquirido por la RNA durante la etapa de entrenamiento, en base a la arquitectura disenada

por medio de la MDRRNA. Utilizando una estrategia de extraccion del conocimiento alma
o mediante programacion una estructura
cenado en los pesos sinapticos
de la red, se disen

similar a la red disenada,

trasladando el conocimiento de la misma a la aplicacion creada.

Como se observa en la figura 3.15, una vez que se entreno y probo la red con los paramet
ros seleccionados por medio de la MDRRNA, se extrajo la matriz de pesos sinapticos
de la
misma. Con esta informacion se creo bajo el entorno de programacion de LabVIEW [42-44] el
esqueleto de esta arquitectura, donde se depositaron los datos correspondientes a la matriz

de pesos de la red disenada

previamente. A la izquierda de la figura, se observa la topologa

problema particular, a esta topologa se le extrae

de una RNA disenada
para resolver algun

el conocimiento almacenado en las matrices de pesos sinapticos

de las diversas capas que la
componen, mostrado al centro de la figura. Con base a la informacion extrada, a la derecha de
la figura se observa que se crea el esqueleto de red, donde se colocan los espacios correspondientes a las diversas capas de neuronas y neuronas dentro de e stas, insertando posteriormente

en ellas los valores extrados de las matrices de pesos sinapticos

de la red disenada.

de herramientas
Figura 3.15: Estrategia de diseno

La caratula
principal del programa NSDann, creado originalmente para un SEEB con 7
esferas de polietileno y un detector de 6 LiI (Eu) para espectros de neutrones expresados en 31

3. Materiales y metodos

grupos de energa se muestra en la figura 3.16. La herramienta NSDann esta dividida en tres
secciones principales. En la parte superior izquierda de la pantalla, se encuentra la seccion
de datos de entrada. En la parte inferior izquierda se localizan los botones de comandos, que
sirven entre otras cosas, para iniciar el proceso de reconstruccion, almacenar la informacion
obtenida y salir del programa. En la parte media se encuentra la seccion de resultados que

muestra de forma grafica

el espectro reconstruido y las dosis calculadas y finalmente, en la
parte derecha de la pantalla se muestran los valores numericos de los correspondientes espectros y dosis calculados.

Figura 3.16: Codigo NSDann6LiI(Eu) para espectrometra neutronica con RNA

La naturaleza intrnseca de las RNA requiere que los datos de entrada, en este caso las
tasas de conteo ledas con el SEEB, se encuentren normalizados a uno para poder realizar la
reconstruccion de los espectros. Debido a e sta naturaleza de las redes, la salida de los datos los

entrega normalizados a un valor de uno tambien. Esta

es una situacion difcil ya que en este
formato, los espectros y dosis obtenidos con las RNA no tienen significado fsico, por lo que no
se pueden aplicar en situaciones experimentales reales debido a que no se pueden interpretar
ni analizar los resultados.

Debido a que el codigo NSDann es una herramienta basada en la tecnologa de RNA, al

momento de crearla se presento el problema expuesto en el parrafo

anterior. La situacion se
una rutina de programacion para normalizar las tasas de conteo en la etaresolvio al disenar

3. Materiales y metodos

pa inicial, y otra rutina que desnormaliza los espectros de neutrones y dosis calculados. Esta

medida hizo practico

el uso de e sta herramienta en ambientes experimentales reales, aunado
a su facilidad y rapidez de operacion. Por e sta razon, los valores numericos de las tasas de con ser escritas
teo provenientes del SEEB, al momento de crear el archivo de entrada, deberan
sin normalizar.

Una de las caractersticas que hacen interesante a e sta herramienta, es el hecho de que

las tasas de conteo ledas con el SEEB, es la unica

informacion que el usuario debe proporcionar al programa, a diferencia de los codigos tradicionales en los que para su ejecucion,

requieren un conjunto de parametros

muy especficos que solo una persona muy especializada es capaz de conocer. Para ingresar las lecturas tomadas con el SEEB es necesario que el
usuario previamente cree un archivo de texto como el mostrado en la figura 3.17.

Figura 3.17: Archivo de tasas de conteo medidas con el SEEB para ingresar a NSDann
En esta figura se puede apreciar que las lecturas de conteo medidas con el SEEB deberan ser colocadas en forma de columna, que en este caso corresponden a las 7 lecturas
tomadas con el SEEB basado en un detector de 6 LiI (Eu), ordenando los datos de la siguiente
manera: en el primer renglon se coloca la lectura tomada con el detector desnudo, y desde el
segundo hasta el septimo renglon se colocan las lecturas tomadas con las esferas de 2, 3, 5, 8,
10 y 12 pulgadas respectivamente.

El usuario cuenta con la capacidad de poder colocar mas de una lectura dentro de mismo
archivo sin restriccion alguna, a excepcion de separar las columnas por medio de tabulaciones.
o de tal forma que se puede seleccionar la columna de datos a ingresar
La herramienta se disen

al programa, escribiendo simplemente el numero

de columna que se desea leer del archivo de

3. Materiales y metodos

entrada a partir del valor 0, que correspondera al primer grupo de tasas de conteo ubicadas a
la izquierda del archivo. Para seleccionar la columna de tasas de conteo que se desea sea leda
por el codigo NSDann, basta con hacer clic en la herramienta Select BSS from file.

De la figura 3.16, se puede observar que para ingresar los valores de entrada correspondientes a las tasas de conteo medidas con el SEEB, el usuario debe dar clic sobre el icono que
tiene la forma de un folder en la seccion BSS count rates. Hecho lo anterior, se abrira una
ventana de seleccion, como la que se muestra en la figura 3.18, que permitira al usuario final
seleccionar la ruta donde se encuentra el archivo que contenga las lecturas tomadas con el
SEEB. Una vez que se ingresan las lecturas de las tasas de conteo medidas con el SEEB, e stas
se muestran en la seccion Spectra BSS Count rate.

Figura 3.18: Seleccion del archivo de entrada con las lecturas BSS al codigo NSDann
Cada vez que se desee reconstruir un espectro diferente, sera necesario seleccionar la
columna de tasas de conteo que se desee ingresar al programa para realizar la reconstruccion
del espectro y cargar el archivo de entrada.

Habiendo ingresado los valores de las tasas de conteo, se esta en condiciones de iniciar el

proceso de reconstruccion del espectro y el calculo

de las 13 dosis que incluyen la dosis equivalente ambiental, H*(10), la dosis equivalente personal, Hp, y la dosis Efectiva, E, a traves de
la RNA de NSDann. Para lo anterior, se da clic en el boton Spectra Unfold. Despues de realizar e sta accion , las tasas de conteo seleccionadas ingresan a la capa de entrada de la RNA

3. Materiales y metodos

embebida en NSDann, iniciandose

la operacion de propagacion de datos por la misma hasta

obtener (en cuestion de segundos) los resultados de salida correspondientes. Esta

informacion

se muestran en forma de graficas

en la parte media de la pantalla, correspondiendo la grafica
superior al espectro de neutrones y la inferior a las 13 dosis equivalentes. Asimismo, en la
parte derecha de la pantalla se muestra la informacion numerica del resultado obtenido, la
columna de la izquierda corresponde al espectro de neutrones y la columna de la derecha a las
dosis equivalentes. Esta informacion se puede almacenar en un archivo tipo reporte, para su

posterior analisis,
tratamiento e impresion.

de la herramienta denominaDebido a los resultados alentadores obtenidos con el diseno

da NSDann para el SEEB con detector de neutrones termicos de 6 LiI (Eu), se repitio el proceso
de una arquitectura de RNA en base a un SEEB con 12 esferas de polietileno y un
de diseno
detector de neutrones termicos de 3 He, para espectros de neutrones expresados en 72 grupos
de energa, como se ilustra en la figura 3.19 .

Figura 3.19: Codigo NSDann3He para espectrometra neutronica con RNA

El principio de operacion del codigo NSDann3He es exactamente el mismo que el descrito

3. Materiales y metodos

en e sta seccion para el codigo NSDann6LiI.

A diferencia de las herramientas de computo tradicionales basadas en codigos de reconstruccion iterativos, y las presentadas en el presente proyecto de investigacion, la deconvolu
cion del espectro de neutrones y el calculo
de las 13 cantidades dosimetricas, a partir del codigo
de reconstruccion de espectros de neutrones NSDann, basado en la tecnologa de RNA, se re
y eficiente. A diferencia de los codigos tradicionales empleados en
aliza de forma rapida,
facil
e sta disciplina, no existe la imperiosa necesidad de proporcionar al programa informacion a

priori o parametro,
a excepcion de las tasas de conteo tomadas a partir de un SEEB. EL resultado se obtiene en cuestion de segundos, pudiendo almacenarse los valores calculados para

un analisis
posterior y continuar de forma inmediata con la reconstruccion de un numero
ilim e sta herramienta resulta
itado de espectros de neutrones. Por sus caractersticas de diseno,

practica
en ambientes experimentales reales.

3.4.3.

NSDEann

o una arquitectura de RNA para un SEEB con detecEn cuestion de segundos se disen
tor de 6 LiI (Eu) empleando el programa NGO. Sin embargo, debe mencionarse que debido a
que NGO es un programa comercial, su codigo fuente no esta disponible, por lo que no es
posible realizar modificaciones al mismo con el objeto de adaptarlo al problema particular de
no ha sido posible extraer el conocimiento
la espectrometra de neutrones, razon por la cual

almacenado en los pesos sinapticos,

lo que representa un serio inconveniente. De igual forma NGO, al ser un programa de proposito general, no esta adaptado para ser aplicado en el
dominio de la espectrometra de neutrones. Como consecuencia, al aplicar NGO de manera
independiente, no es posible analizar e interpretar los resultados obtenidos. La principal dificultad, es que al tratarse de RNA, los datos de entrada y salida se encuentran normalizados.
Aunado a lo anterior, NGO no es capaz de presentar los espectros de neutrones y dosis en el

formato establecido por los o rganos reguladores en este ambito

del conocimiento. Todas e stas
dificultades hicieron evidente la necesidad de crear una herramienta tecnologa que resolviera
los problemas mencionados.

Por la razon anterior, se creo la herramienta de computo denominada NSDEann6LiI

3. Materiales y metodos

(de sus siglas en ingles, Neutron Spectrometry and Dosimetry by means of Evolutinary Artificial Neural Netoworks), mostrada en la figura 3.20, para un SEEB basado en un detector
de 6 LIi(Eu). Como puede observarse en e sta figura, el aspecto y principio de operacion es muy
similar al descrito para el codigo NSDann.

Figura 3.20: Codigo de reconstruccion de espectros de neutrones NSDEann6LiI

Debido a las limitaciones impuestas por el programa NGO, el procedimiento de recon
struccion de los espectros de neutrones el y calculo
de las 13 dosis equivalentes se realiza a
traves de las siguientes etapas. En primer lugar, es indispensable tener instalado el programa
NGO en la computadora donde se ejecute el codigo NSDEann, ya que ambos programas trabajan en estrecha relacion al ser interdependientes.

Una vez en ejecucion los programas NGO-NSDEann, es necesario cargar en NGO la

RNAE disenada
con este programa, para el SEEB que se desee realizar la reconstruccion del
espectro.

Cabe mencionar que al trabajar con RNA, por la naturaleza intrnseca de las mismas, es
requisito que los datos de entrada a la red esten normalizados a un valor de uno. NGO no es
la excepcion y para realizar la reconstruccion del espectro, es requisito que las tasas de conteo
ledas con el SEEB esten normalizadas a este valor, por lo que fue necesario crear una rutina

3. Materiales y metodos

de normalizado de las tasas de conteo en el codigo NSDEann. De forma similar y a causa de las
caractersticas naturales de las RNA, los espectros de salida obtenidos con NGO, tambien se

encuentran normalizados a uno, lo que en el ambito

de las espectrometra de neutrones carece
de significado fsico, razon por la cual fue necesario crear en NSDEann una rutina para el

desnormalizado de los espectros y dosis calculados, con el objetivo de poder realizar un analisis de los datos.

La siguiente etapa consiste en introducir las tasas de conteo normalizadas a la RNAE

creada con NGO. Para ello se utiliza la herramienta de normalizacion de NSDEann. Para la
normalizacion de las tasas de conteo, se carga el archivo de datos de entrada sin normalizar,
correspondiente a las lecturas del SEEB, de la misma forma como fue descrito en la seccion del
codigo NSDann. Posteriormente, al presionar el boton Saved normalized counts, se abre una
ventana de seleccion como la mostrada en la figura 3.21, que permite almacenar la informacion

de las tasas de conteo normalizadas en un archivo de texto. Esta

informacion se va a utilizar
en el programa NGO para reconstruir el espectro de neutrones y calcular las 13 dosis.

Figura 3.21: Archivo de tasas de conteo sin normalizar para NSDEann

Las tasas de conteo normalizadas se ingresan a la herramienta de prediccion de NGO,
mostrada en la figura 3.22, y se reconstruye el espectro de neutrones y las 13 cantidades

3. Materiales y metodos

dosimetricas, que como se ha mencionado previamente se encuentran normalizadas a uno.

Figura 3.22: Herramienta de prediccion de NGO para espectrometra neutronica

La informacion correspondiente al espectro de neutrones y dosis equivalentes, como se
observa en la figura 3.23, se almacena en una archivo de texto de forma similar al procedimiento descrito para crear el archivo de entrada de las tasas de conteo para el codigo NSDann, esto
es, el espectro y dosis se almacenan en una columna, correspondiendo los primeros renglones

al espectro de neutrones y los ultimos

13 a las cantidades dosimetricas. No existe restriccion

en cuanto al numero
de espectros y dosis que se puedan almacenar en este archivo, excepto
que se encuentren separados por tabulaciones.

de NSDEann es muy similar al diseno

del codigo NSDann, por lo que para poder
El diseno

analizar los espectros de neutrones obtenidos con la red disenada

mediante NGO, el archivo
creado en la etapa anterior se carga a traves de la herramienta Evolved Spectra&Doses. Al
dar un clic en el icono en forma de carpeta, se abre una ventana de seleccion en donde puede
buscarse la ruta donde se encuentra almacenado el archivo de texto que contiene la informa
cion de los espectros y dosis que se desea analizar. Para mostrar la grafica
y los valores de los
espectros y dosis desnormalizados, se da clic al boton Plot Spectra&Doses.

Al igual que el codigo NSDann, en la parte inferior izquierda de la pantalla de NSDEann

se tiene una serie de botones que permiten entre otras cosas, iniciar el proceso de reconstruccion a la RNA, almacenar los valores obtenidos en este proceso en un archivo de texto, para un

analisis
posterior y, salir del programa.

3. Materiales y metodos

normalizados a uno
Figura 3.23: Archivo con espectros reconstruidos con NGO que estan
Debido a las bondades de programacion de LabVIEW, fue posible modificar el codigo

NSDEann6Li para aplicarlo a una RNAE disenada

para un SEEB basado en un detector de
3 He,

como se aprecia en la figura 3.24.

Figura 3.24: Codigo de reconstruccion de espectros de neutrones NSDEann3He

El principio de operacion de este codigo es el mismo que el descrito para NSDEann6Li.

Debido a que la herramienta desarrollada y explorada en este trabajo, se encuentra en

se observa que aun
existen multiples

una etapa temprana de diseno,

inconvenientes para

3. Materiales y metodos

aplicarla de forma eficiente, como por ejemplo la dependencia con un programa externo que
es comercial, la seleccion de los parametros

ademas
de configuracion de los AE, la incapacidad
de presentar la informacion obtenida de acuerdo con las normas establecidas para la espec no existen reglas fijas para
trometra neutronica, entre otras. Sin embargo, se observa que aun
resolver los problemas antes mencionados, por lo cual es necesario realizar mas investigacion
con el objetivo de construir las teoras y herramientas que permitan resolver los problemas
planteados. Lo anterior abre una linea de investigacion en donde se manifiesta la necesidad de
desarrollar herramientas a la medida, con el objeto de aplicar de forma eficiente la tecnologa
realizando trabajos tendientes a proponer soluciones a
analizada. En la actualidad se estan
los problemas encontrados.

3.4.4.

NSDTB

de RNA apliEn el presente trabajo se exploraron diversas metodologas para el diseno

cadas en el dominio de la espectrometra neutronica. Para cada red disenada

se creo una
en ambientes experimenherramienta de computo con el objetivo de observar su desempeno
tales reales. Debido a lo anterior, y como se aprecia en la figura 3.25, se presento la necesidad
de intercomparar los resultados obtenidos con las herramientas creadas, as como contra los
metodos tradicionales empleados en la actualidad. De igual forma se presento la necesidad de
de cada herramienta con relacion a las demas.

evaluar el desempeno

Por la razon anteriormente expuesta se ilustra en la figura 3.25. Para resolver el prob se creo la herramienta denominada NSDTB6Li (de
lema, en una primera etapa de diseno
sus siglas en ingles, Neutron Spectrometry and Dosimetry Tool Box), mostrada en la figura
3.26, para un SEEB compuesto a base de un detector de 6 LiI (Eu), la cual tiene como proposito
fundamental comparar los espectros de neutrones y dosis calculadas, obtenidos con diversas
metodologas de reconstruccion de espectros de neutrones. Dicha herramienta tiene la capacidad de realizar la prueba estadstica de la Chi cuadrada entre los espectros y dosis calculados
con las diversas metodologas. Lo anterior permite comparar estadsticamente la informacion
citada.

3. Materiales y metodos

Figura 3.25: Necesidad de herramienta para comparar diversas metodologas

del programa es muy semejante al de los codigos NSDann y NSEl formato de diseno
DEann, como se ilustra en la figura 3.26.

Figura 3.26: NSDTB:compara espectros obtenidos con diversas metodologas

Para comparar espectros y dosis calculados con diferentes metodologas de deconvolucion, lo primero es crear un archivo de texto, de forma semejante a como se explico en la
seccion 3.4.2.

ser colocados en forma de columna, como se ilustra en la figuLos espectros deberan

3. Materiales y metodos

ra 3.27, en donde se puede apreciar que los primeros renglones corresponden al espectro de

neutrones y los 13 ultimos

a las dosis equivalentes.

Figura 3.27: Archivo con espectros y dosis a comparar con el codigo NSDTB6Li
para comparar hasta cinco espectros de neuEsta herramienta de computo se diseno
trones y dosis equivalentes calculados con diferentes codigos de deconvolucion. La operacion
de NSDTB es la siguiente. Despues de haber creado el archivo de texto con los espectros y
dosis a ser comparados, el usuario debe dar clic en el icono en forma de carpeta en la herramienta File of spectra and doses to compare. Enseguida se abre una ventana de seleccion
desde donde el usuario puede elegir la ruta donde se encuentra almacenado este archivo. Posteriormente, solo es necesario dar clic al boton Plot Spectra&Doses y espectros y dosis se
mostraran en la parte media del programa.

Esta
herramienta de computo fue disenada
para comparar espectros y dosis por medio
o una rutina de programacion que aplica la
de la prueba estadstica Chi cuadrada. Se disen
En la parte derecha de la
prueba Ci cuadrada a cada espectro con relacion a todos los demas.
pantalla del codigo NSDTB puede apreciarse un arreglo de datos que muestra los resultados
al aplicar e sta prueba.

Al comparar las herramientas clasicas

con las presentadas en el presente trabajo de
no disponen de una herinvestigacion, se observo que ambas carecen de un elemento comun:

3. Materiales y metodos

ramienta que permita cuantificar la incertidumbre asociado con el proceso de medicion. Este
aspecto es importante, ya que la naturaleza propia del proceso de reconstruccion del espectro
de neutrones mediante el empleo del SEEB, se ve afectado por varias incertidumbres, entre las
cuales se puede mencionar: el proceso de lectura de las tasas de conteo con el SEEB, el metodo
y la calidad con la que fue calculada la matriz de respuesta del sistema, la calidad y eficiencia
del detector de neutrones. En vista de la necesidad de cuantificar la incertidumbre asociada

con el proceso de medicion en el ambito

de la espectrometra de neutrones, se analizo el uso de
la estadstica de Poisson, tratando de proponer una alternativa al problema enfrentado.

A continuacion se describe brevemente el desarrollo de una rutina de programacion, que

incorporada al codigo NSDann permite cuantificar a traves de la estadstica de Poisson, la
incertidumbre asociada con el proceso de medicion con el SEEB.

3.5.

de las herramientas disenadas

Optimizacion
utilizando la
estadstica de Poisson
La distribucion de Poisson juega un papel de fundamental importancia en el trabajo apli-

cado para modelar problemas de conteo en muchas areas

[245248]. Se se llama as en honor

a Simeon Dennis Poisson (1781-1840), frances que desarrollo esta distribucion basandose
en

estudios efectuados en la ultima

parte de su vida. Esta
distribucion es una distribucion de

probabilidad discreta que expresa la probabilidad de un numero

de eventos.

La distribucion de Poisson se utiliza cuando se hacen registros de eventos que se distribuyen al azar en un espacio o tiempo determinado. Puede esperarse que cierto proceso obedezca la ley de Poisson y ante esta suposicion se puede calcular la probabilidad de que ese

evento se presente en una unidad de tiempo, por ejemplo, el numero

de llegadas a un super

mercado por minuto, el numero

de defectos en piezas similares de un material, el numero
de

errores de ortografa cometidos al escribir en una pagina,

el numero
de estrellas en un deter

minado volumen de espacio, el numero

de goles que marca un equipo en un partido de futbol,

el numero
de neutrones que entran al SEEB cada segundo, etc.

3. Materiales y metodos

En una fuente de neutrones, se lleva a cabo una secuencia de eventos aleatorios independientes, es decir, que la ocurrencia de cualquier evento no tiene efecto en la ocurrencia de
cualquier otro. Para obtener el numero de emisiones que genera una fuente radiactiva, que
en nuestro caso se centra en una fuente de neutrones, se utiliza el SEEB con un detector del

tipo 6 LiI (Eu) o 3 He. Basicamente,

el funcionamiento de los detectores consiste en convertir a
un pulso electrico cada neutron detectado. Este pulso electrico se amplifica y se transmite a un
sistema multicanal, controlado a traves de un programa de PC, con el cual se puede realizar

la adquisicion de las tasas de conteo, esto es, el numero

de veces que el detector recibe una
sin tener conocimiento de la intensidad de la misma. Es posible realizar este conteo en
senal
distintos intervalos de tiempo y cierta cantidad de veces.

En el estudio del comportamiento estadstico del decaimiento radiactivo de una fuente

de neutrones, se propone la distribucion de Poisson a partir de los datos experimentales
obtenidos, esto es, las tasas de conteo. El objetivo es obtener una funcion de distribucion definida por una variable aleatoria, denotada por n, la cual se considera en el siguiente desarrollo

como el numero
de emisiones de la fuente radiactiva de neutrones, cuya caracterstica es que
es discreta y no negativa. Con e sta distribucion se pretende cuantificar la incertidumbre asociada con la medicion de las tasas de conteo ledas con el SEEB.

La distribucion de Poisson se refiere a ciertos procesos que pueden ser descritos con una
asumir
variable aleatoria discreta. La letra X suele representar esa variable y puede ademas

valores enteros (0,1,2,3 etc..). Se utiliza la letra X mayuscula

para representar la variable

aleatoria y la x minuscula
para designar un valor especfico que puede asumir la X mayuscula. La probabilidad de exactamente x ocurrencias en una distribucion de Poisson se calcula
mediante la formula:
P ( k ) = bk

eb
k!

(3.4)

donde b = a 1 y k = 1, 2, ..., a es el parametro

de una variable aleatoria exponencial y la
variable aleatoria de Poisson es el entero k.

Como se aprecia en la figura 3.28, la entrada al sistema son las tasas de conteo prove-

3. Materiales y metodos

nientes del SEEB. Posteriormente e stos datos ingresan a la rutina Poisson, lo que genera una
matriz de datos de entrada alrededor del valor central (valor experimental). La matriz de
tasas de conteo compuesta por las experimentales y las calculadas con la distribucion Poisson,
se introducen una a la vez a la rutina principal del programa NSDann.

(a) Tasas conteo (b) Matriz de tasas de conteo producidas por la distribucion Poission,
Experimentales alrededor de los valores experimentales

Figura 3.28: Distribucion de Poisson en tasas de conteo ledas con el SEEB

Cuando todas las tasas de conteo pasan a traves de RNA, se genera un conjunto de
espectros de de neutrones y dosis equivalentes, como se ilustra en la figura 3.29.

(a) Espectros alrededor del valor experimental

(b) Dosis alrededor del valor experimental

Figura 3.29: Espectros y dosis calculados con NSDann, alrededor de la distribucion de Poisson

Finalmente se aplica la media aritmetica y la desviacion estandar

para obtener el es-

3. Materiales y metodos

pectro final y 13 dosis equivalentes con el error asociados a cada uno, como se muestra en la
figura 3.30.

(a) Espectros e incertidumbres

(b) Dosis e incertidumbres

Figura 3.30: Espectros y dosis con incertidumbres, calculadas con la distribucion de Poisson

Es importante mencionar que la estadstica de Poisson se aplico para medir una sola
fuente de incertidumbre: las tasas de conteo medidas con el SEEB, sin embargo, la espectrometra neutronica por medio del SEEB se ve afectada por otras incertidumbres como la
asociada con la matriz de respuesta, la que tiene que ver con la instrumentacion electronica o
la relacionada con los codigos de deconvolucion utilizados para reconstruir el espectro. Sera
deseable cuantificar las incertidumbres antes mencionadas, por lo que la se podra utilizar la
distribucion normal con este proposito. Actualmente se realizan estudios para incluir en los
espectros de neutrones calculados por las herramientas aqu presentadas las incertidumbres
antes mencionadas.

3.6.

de las herramientas en experimentos reales y

Aplicacion
con tecnicas

comparacion
clasicas
Construidos los prototipos de las herramientas computacionales, la etapa siguiente con-

Por ello,
sistio en aplicarlas en situaciones experimentales reales, para evaluar su desempeno.

en e sta etapa, se se utilizaron las herramientas disenadas

el el proceso de reconstruccion de

los espectros de neutrones y el calculo

de dosis, a partir de resultados experimentales reales,
traducidos en las tasas de conteo medidas con el SEEB ante fuentes de neutrones.

3. Materiales y metodos

En el procedimiento experimental, se utilizaron las tasas de conteo obtenidas en diversos experimentos realizados con los SEEB de la Unidad Academica de Estudios Nucleares
(UAEN), perteneciente a la Universidad Autonoma de Zacatecas (UAZ), Mexico, del Departa
mento de Ingeniera Nuclear (DIN), de la Universidad Politecnica de Madrid (UPM) en Espana
y del Laboratorio de Patrones Neutronicos (LPN), perteneciente al Laboratorio de Metrologa
de Radiaciones Ionizantes (LMRI) del Centro de Investigaciones Energeticas Medio Ambien
tales y Tecnologicas (CIEMAT), con sede en Madrid, Espana.

El SEEB de la UAZ y UPM, como se ilustra en la figura 3.31, esta basado en un detector
de neutrones termicos compuesto de Ioduro de litio activado por Europio 6 LiI (Eu), con 7 esferas
de polietileno cuyas dimensiones son 0, 2, 3, 5, 8, 10 y 12 pulgadas respectivamente, con una
matriz de respuesta conocida como UTA4, expresada en 31 grupos de energa.

Figura 3.31: SEEB con detector 6 LiI (Eu)

El SEEB utilizado por el CIEMAT, como se observa en la figura 3.32, esta basado en un
detector de neutrones termicos compuesto de Helio-3 3 He, con 12 esferas de polietileno cuyas
dimensiones son 3, 3.5, 4, 4.5, 5, 6, 7, 8, 9, 9.5, 10 y 12 pulgadas respectivamente, con una
matriz de respuesta expresada en 72 grupos de energa.

3. Materiales y metodos

(a) Detector 3 He

(b) Esferas Bonner 3 He

Figura 3.32: SEEB con detector 3 He

Cuando se utiliza el SEEB, como el mostrado en la figura 3.33, se miden las tasas de
conteo que los neutrones producen en cada una de las configuraciones detector-moderador.
La tecnica de las esferas Bonner, involucra recolectar pulsos electronicos, a traves de la elec
tronica asociada al SEEB, que surgen de las interacciones entre los neutrones y los atomos
de
6 Li

por ejemplo, en un detector de Ioduro de Litio dopado con Europio, 6 LiI (Eu), con el detector

en diferentes configuraciones de conteo en el campo de irradiacion de neutrones.

Figura 3.33: Electronica asociada del SEEB para medir las tasas de conteo
Para obtener las tasas de conteo mediante el SEEB, se realiza una serie de medidas a
cierta distancia respecto de la fuente de neutrones, colocando el detector de neutrones en el

3. Materiales y metodos

centro de cada esfera de polietileno. La combinacion detector-moderador tendra una respuesta

diferente a los neutrones en funcion de la energa. Los neutrones rapidos

se frenan dentro
del moderador y llegan al detector en estado termico, mientras que los neutrones inicialmente
termalizados son parcialmente capturados dentro del moderador y no alcanzan el detector. Au
mentando el diametro
de la esfera moderadora, se desplaza el pico de sensibilidad del sistema
haca las energas altas y de esta manera es posible efectuar una espectrometra de neutrones.

Tanto para la espectrometra como para la dosimetra de neutrones, se utilizan estas interacciones, ya que al producir partculas cargadas, e stas inducen en la instrumentacion por

medio de la ionizacion, senales

electricas a partir de las cuales y mediante procedimientos adecuados, se remonta a la energa del neutron detectado. A partir de la tasas de conteo tomadas
proporciona informacion
con las esferas, es posible reconstruir el espectro de neutrones, el cual
acerca de la distribucion de energa de los neutrones incidentes. El simple hecho de medir las
tasas de conteo con el SEEB, no permite realizar la reconstruccion del espectro.

Para ello es necesario hacer uso de diversos codigos de deconvolucion que resuelvan la
ecuacion integral diferencial de Fredholm de primer tipo, que como se ha comentado en sec
ciones anteriores, es un sistema de ecuaciones mal condicionado, con un numero
infinito de
soluciones. El seleccionar aquella solucion que tenga significado para el tipo de problema, es
procedimiento para la
parte del proceso de reconstruccion, por lo que se debe aplicar algun
deconvolucion del espectro.

Al utilizar las cuentas integrales ledas a partir del SEEB es posible reconstruir el espectro de neutrones empleando varios metodos, entre los que se pueden mencionar: los metodos

Monte Carlo, de la Maxima

Entropa, de Parametrizacion e Iterativos. Sin embargo, como
puede observarse en la figura 3.34, dichos metodos se ejecutan en ambientes de consola de
comandos como MS-DOS, lo que hace muy complicado su uso. En el caso de la UAZ y la UPM
se utiliza el codigo conocido como BUNKIUT, un codigo basado en metodos iterativos, desarrollado en la Universidad de Austin en Texas, Estados Unidos. Para el caso del CIEMAT, se
utiliza el codigo conocido como Unfolding with Maxel and Gravel (UMG), un codigo iterativo

basado en el principio de la Maxima

Entropa, desarrollado en el PTB de Alemania.

3. Materiales y metodos

Figura 3.34: Ambiente de codigos tradicionales para espectrometra neutronica

del inconveniente mencionado en el parrafo

Ademas
anterior, las principales dificultades
de e stos metodos son: la alta dependencia de la experiencia del experimentador y de la calidad
de la informacion a priori (espectro inicial), que debe ser proporcionada a los mismos.

reEn la figura 3.35, se puede observar en entorno de trabajo del codigo UMG, el cual

quiere una serie de archivos y parametros

muy especializados antes de realizar el proceso de
reconstruccion de los espectros.

Figura 3.35: Ambiente de trabajo del codigo UMG

De la figura 3.35, se puede observar que para la operacion de este codigo, el cual se eje-

3. Materiales y metodos

cuta en el ambiente de trabajo MS-DOS, se requiere un conjunto de archivos con informacion

muy especfica. A continuacion se describe brevemente el cada archivo:
Se requiere un archivo de entrada.Este archivo sirve como referencia para un conjunto
de archivos necesarios en el ambiente UMG para realizar la deconvolucion. El archivo de
entrada recibe el nombre de: gv f cC IE.inp para la deconvolucion con el programa GRAVEL
y mx f cC IE.inp cuando se usa el codigo MAXED. En este archivo se indican los nombres
de los restantes archivos de entrada necesarios
Se necesita un archivo de medidas (.ibu). Se requiere hacer saber al programa el archivo
que contenga las tasas de conteo medidas con el SEEB.
Es necesario un archivo con matriz respuesta (. f mt). Es necesario hacerle saber al programa el archivo que contiene la matriz de respuesta del sistema espectrometrico empleado.
Es importante declarar el nombre del archivo de salida ().
Es indispensable y de fundamental importancia un archivo con los datos del espectro
inicial (. f lu). Para realizar el proceso de deconvolucion con este codigo, es necesario
parecido al
emplear un espectro inicial. Debido a que el espectro inicial debe ser lo mas
problema que se desea resolver, y que en la mayora de los casos se desconoce su solucion,
esto representa un serio inconveniente de este tipo de herramientas.
Cabe resaltar que para un correcto funcionamiento del codigo UMG, es necesario respetar el formato adecuado para cada uno de los archivos de entrada. De no respetarse, el

programa no funcionara.

En el proceso de creacion de cada uno de los archivos de entrada mencionados previa

mente, es necesario establecer una serie de parametros

para poder llevar a cabo el calculo.

Mediante un proceso de ensayo y error, e stos parametros

se varan en la busqueda
de una
solucion adecuada.

Al haber obtenido el espectro de neutrones, las dosis equivalentes se calculan por medio
de los coeficientes de fluencia de dosis, que en la mayora de los codigos de reconstruccion, son

3. Materiales y metodos

coeficientes no actualizados lo que conlleva a subestimarla, o bien, a obtener una dosis que ya
no se utiliza.

En los metodos tradicionales, como en el caso de los codigos BUNKIUT y UMG, la informacion de los espectros se utiliza para calcular las magnitudes dosimetricas utilizando los
coeficientes del ICRP 74 (ICRP 1996). Como se aprecia en la figura 3.36, en el caso de las

herramientas disenadas
en el presente trabajo, por ejemplo el codigo NSDann, los espectros

de neutrones y 13 magnitudes dosimetricas son calculados de forma simultanea

e indepen

diente por la RNA en un ambiente grafico,

amigable, intuitivo, facil,
rapido
e independiente
de cualquier programa de computo, utilizando tan solo como informacion de entrada las tasas
de conteo medidas con el SEEB. Asimismo este codigo muestra la incertidumbre asociada con
el proceso de medicion de las tasas de conteo ledas con el SEEB, a diferencia de los codigos

tradicionales que no efectuan

dicho calculo.

(a) Espectro de neutrones con incertidumbres

(b) Dosis equivalentes con incertidumbre

Figura 3.36: Espectros y dosis calculados con el codigo NSDAnn y sus incertidumbres

Captulo 4

Resultados
Resultados parciales de las tecnicas y herramientas mostradas en el presente proyecto

de investigacion, fueron presentados en multiples

congresos internacionales. Los resultados
relevantes se publicaron en revistas, proccedings y un captulo de libro. Los productos
mas
derivados se enlistan a continuacion.

4.1.

Revistas

1. Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. NeuFillol A., Mendez-Villafane
tron spectrometry using artificial neutral networks for a Bonner sphere spectrometer with a He-3 detector. Revista Mexicana de Fsica, 57(1):69-71, 2011. (Sociedad Mexicana de Fsica, Mexico, ISSN: 0035-00IX(print)).

de una RNA para el sistema especEn este documento se describe el proceso de diseno
En el protrometrico de esferas Bonner perteneciente al LMRI del CIEMAT en Espana.
de la arquitectura o ptima se empleo la MDRRNA. En un trabajo adicional,
ceso de diseno
la arquitectura de otra RNA empleando
que no se incluye en este documento, se diseno
la metodologa de RNAE.

2. Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. PerFillol A., Mendez-Villafane
formance of artificial neutral networks and genetical evolved artificial neutral

4. Resultados

networks unfolding techniques. Revista Mexicana de Fsica, 57(1):89-92, 2011. (Sociedad Mexicana de Fsica, Mexico, ISSN: 0035-00IX(print)).

Se compara el rendimiento y capacidad de generalizacion de RNA disenadas

para el
con dos metodologas de diseno:
MDRRNA y RNAE.
SEEB del LMRI-CIEMAT, Espana,

3. Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A neutron spectrometry and dosimetry computer tool based on ANN. 12th International Congress of the International Radiation Protection Association (IRPA12). IAEA Procceding Series, STI/PUB/1460, 1-9. (International Atomic Energy Agency, Vienna, ISBN
978-92-0-105410-4).

En este trabajo se realiza el desarrollo de una herramienta de computo en el lenguaje

de programacion de Matlab, que permite comparar espectros de neutrones con varias
metodologas de reconstruccion de espectros. Se observo la necesidad de utilizar una
plataforma de programacion diferente, ya que se presentaron inconvenientes en el uso

practico
de la herramienta, sobre todo en ambientes experimentales reales.

4. Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Neutron spectrometry and dosimetry based on a new approach called Genetic Artificial Neural Networks. 12th International Congress of the International Radiation
Protection Association (IRPA12). IAEA Procceding Series, STI/PUB/1460, 1-9. (International Atomic Energy Agency, Vienna, ISBN 978-92-0-105410-4).

Se realiza un estudio inicial de la interaccion entre RNA y AE. Se observo que el diseno
de RNA por medio de AE, utilizando la herramienta conocida como NGO, se realizaba en
un periodo de tiempo muy inferior al compararlo con las tecnicas tradicionales (ensayo

resulta ser un serio problema con

y error)), sin embargo, la seleccion de parametros
aun
e sta metodologa, difcil de resolver.

4. Resultados

5. Vega-Carrillo H.R., Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Hernandez-D

avila
V.M. ANN in spectrometry and neutron dosimetry. American Institute of Physics
Proccedings, 1310:1217, 2010.

Se describen las aptitudes de las RNA aplicadas en la espectrometra de neutrones, con

la solucion de ejemplos experimentales reales. En este trabajo, todava no se utilizan las
herramientas tecnologicas presentadas en el presente proyecto de investigacion.

de las arquitecturas de RNA (ensayo

Se utilizan las metodologas clasicas
para el diseno
y error), con sus correspondientes inconvenientes. Los espectros de neutrones se reconstruyen en el entorno de Matlab, lo que hace muy difcil su implementacion en ambientes
experimentales reales. Lo anterior se debe a que se requiere tener instalado el programa
Matlab y, posteriormente se requiere tener el archivo fuente que contenga la arquitec

tura de red disenada

y los pesos sinapticos
de misma.

de lo antes mencionado, es indispensable que para operar e sta herramienta, el

Ademas
usuario cuente con bases solidas en lo que se refiere a la teora de RNA as como en la
operacion general del entorno de Matlab. Para aplicar con e xito e sta tecnologa en el proceso de reconstruccion de espectros de neutrones por medio de e sta tecnica, se requiere
tener conocimiento en el uso de la herramienta especializada que se utiliza en
ademas
Matlab para trabajar con RNA (toolbox de RNA). Lo anterior, para un usuario no experimentado implica un serio inconveniente.

Otro inconveniente con el uso de e sta metodologa, es el hecho de que para realizar el

calculo
de las dosis equivalentes, se requiere utilizar una RNA diferente. Lo anterior implica aplicar la tecnica de ensayo y error para determinar la topologa de red, tecnica que
presenta grandes inconvenientes. Para reconstruir los espectros y calcular las dosis es
indispensable que el usuario cuente con los dos archivos fuente que contengan sendas
redes entrenadas para cada caso particular, y ejecutar en procesos separados cada una
para obtener los resultados buscados.

4. Resultados

de e sta
Un inconveniente adicional se presenta cuando se desea comparar el desempeno

tecnologa contra el obtenido con metodos clasicos.

Como se puede ver, todo lo anteri
or hace evidente que este procedimiento es poco eficiente y presenta multiples
inconve las herramientas tecnologicas presentadas en el presente
nientes, que orillaron a disenar
trabajo.

6. Vega-Carrillo H.R., Ortiz-Rodrguez J.M., Hernandez-D

avila
V.M., Martnez-Blanco M.R.,

B., Ortiz-Hernandez

Hernandez-Almar
az
A.A., and Mercado-Sanchez
G.A. Different
spectra with the same neutron source. Revista Mexicana de Fsica, 51(6):3539,
2010.

En este documento se muestra la capacidad de las RNA en el ambito

de la espectrometra
de neutrones con la solucion de un problema experimental real. En este trabajo, todava

no se hace uso de las herramientas disenadas.

Al igual que en trabajo comentado con

y operacion de las RNA aplianterioridad, aqu se utiliza la tecnica clasica

para el diseno
cadas en la espectrometra neutronica, con sus correspondientes inconvenientes.

7. Ortiz-Rodrguez, J.M.; Martnez-Blanco, M.R.; Gallego, E.; Vega-Carrillo, H.R. Evolutionary Artificial Neural Networks in Neutron Dosimetry. American Nuclear Society Transactions, 101:647-648. (American Nuclear Society, ISSN: 0003-018X(Print)).

de multiples

Se destaca la rapidez de los AE en el diseno

arquitecturas de RNA aplicadas
en la espectrometra neutronica y se presenta una herramienta de computo en el entorno
de programacion de LabVIEW para la comparacion de espectros de neutrones calculados

con las RNAE comparandolos

con las metodologas tradicionales. Se destaca el hecho de

que e sta herramienta permite comparar resultados obtenidos con multiples

codigos de
reconstruccion de espectros.

8. Vega-Carrillo H.R., Martnez-Blanco M.R., Hernandez-D

avila
V.M., and Ortiz-Rodrguez
J.M. Spectra and dose with ANN of

252C f , 241 Am Be,

and

239 Pu Be.

Journal of Ra-

4. Resultados

100

dioanalytical and Nuclear Chemistry, 281(3):615618, 2009.

En este documento se muestra la capacidad de las RNA en el ambito

de la espectrometra
de neutrones con la solucion de un problema experimental real. En este trabajo, no se

hace uso de las herramientas disenadas.

Al igual que los trabajos comentados con anteri
y operacion de las RNA aplicadas
oridad, aqu se utiliza la tecnica clasica
para el diseno
en la espectrometra neutronica, con sus correspondientes inconvenientes.

9. Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A computational tool design for evolutionary artificial neural networks in neutron
spectrometry and dosimetry. Proceedings of the Electronics, Robotics and Automotive
Mechanics Conference (CERMA09), IEEE Computer Society,113118, 2009.

de una herramienta tecnologica que permite comparar los especSe presenta el diseno
tros de neutrones obtenidos con la tecnologa de redes neuro-evolucionadas contra las

obtenidas con tecnicas clasicas.

Pronto se observo que tambien se podan intercomparar

espectros de neutrones reconstruidos con multiples

codigos de deconvolucion.

10. Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann, a LabVIEW tool for neutron spectrometry and dosimetry based on the RDANN methodology. Proceedings of the Electronics, Robotics and Automotive Mechanics Conference
(CERMA09), IEEE Computer Society, 00:131136, 2009.

de un
Resultado de un proyecto de tesis de maestra tutelada, que describe el diseno

codigo para la reconstruccion de espectros de neutrones grafico,

amigable, intuitivo, facil

de operar, rapido
y que fue disenado
en el lenguaje de programacion de LabVIEW. Esta
o en base al conocimiento extrado a una RNA, disenada

herramienta se disen
por medio
de la MDRRNA.

11. Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. A New Computer

4. Resultados

101

Tool Based on ANN Technology, applied in neutron spectrometry and dosimetry research areas. Proceedings of the Electronics, Robotics and Automotive Mechanics
Conference (CERMA08), IEEE Computer Society, 189-194, 2008.

Este trabajo, parte de un proyecto de tesis de maestra tutelada, forma parte de las in una herramienta de computo para
vestigaciones preliminares que permitieron disenar
comparar los espectros de neutrones obtenidos con diversos codigos de reconstruccion.
e sta herramienta se desarrollo dentro del entorno de
En las etapas tempranas de diseno,
programacion de Matlab. Observando resultados satisfactorios y de acuerdo con experimentos realizados, se pudo apreciar que dicha herramienta presentaba inconvenientes
que no la haca apropiada para ser aplicada en situaciones experimentales reales, por
lo que, con la experiencia obtenida, se proyecto migrar a una plataforma de computo
que permitiera desarrollar una version de este programa que si pudiera ser utilizada en
situaciones reales.

12. Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Artificial neural networks modeling evolved genetically, a new approach applied
in neutron spectrometry and dosimetry research areas. Proceedings of the Electronics, Robotics and Automotive Mechanics Conference (CERMA08), IEEE Computer
Society, 387392, 2008.

Se presenta el analisis
realizado al estudiar la interseccion de las RNA con AE, en el
de arquitecturas de RNA mediante e sta tecnologa
proceso de optimizacion de diseno

hbrida en el area
de espectrometra neutronica.

4.2.

Captulo de libro
Ortiz-Rodrguez, J.M.; Martnez-Blanco, M.R. and Vega-Carrillo, H.R. Evolutionary artificial neural networks in neutron spectrometry. Captulo de libro publicado en:
Chi Leung Patrick Hui. (editor) ARTIFICIAL NEURAL NETWORKS - APPLICATION,
Ed. INTECH (2011). [ISBN: 978-953-307-658-4]. In Press.

4. Resultados

102

de arquiEn este documento se analiza con mayor profundidad el uso de AE en el diseno

tecturas de RNA aplicadas en el dominio de la espectrometra de neutrones por medio de
tecnicas de IA.

Se presenta una herramienta disenada

en el entorno de programacion de LabVIEW que
tiene como objetivo fundamental resolver los problemas encontrados al utilizar el software comercial NGO en el modelado de las redes del problema a resolver. Dicha herramienta permite aplicar la tecnologa de RNAE en situaciones experimentales reales,
cosa que no es posible con el simple uso de NGO.

otra herramienta que permite intercomparar y analizar, grafica,

Se presenta ademas
numerica y estadsticamente por medio de la prueba estadstica Chi cuadrada, espectros
de neutrones obtenidos con diversos codigos de reconstruccion de espectros de neutrones.

Se expone la problematica
existente en el problema de determinar los parametros
de
y configuracion de los AE, un problema de optimizacion para el cual, en la actualdiseno
idad, no se cuenta con elementos cientficos y metodologicas que permitan resolverlo de
forma satisfactoria.

A pesar de los problemas existentes en la determinacion de los parametros

de los AE,

Se disenaron
varias arquitecturas de RNAE por medio de la tecnica de ensayo y error y

se observo que los resultados obtenidos con las distintas arquitecturas de red disenadas,
presentan resultados similares en la reconstruccion de espectros de neutrones, lo que
permite ver que es una tecnologa prometedora en la solucion de este tipo de problemas.

El buscar y proponer soluciones al problema de optimizacion no lineal en la determi

nacion de los parametros

de AE, abre una linea de investigacion en la busqueda
de re
solver este tipo de problemas en e sta area
de conocimiento particular.
Enseguida se presentan, los documentos previamente comentados.

University Campus, STeP Ri
Slavka Krautzeka 83/A
51000 Rijeka, Croatia

T/F +385 51 686 166
E info@intechweb.org
www.intechweb.org

CERTIFICATE OF INVITATION
June 01, 2010

To Whom it May Concern,

InTech is an Open Access publisher of scientific books and journals. We strive to produce high quality
publications that are of benefit to the community of scholars throughout the world. Thus, when selecting
authors to contribute to our publications we apply a rigorous set of criteria. InTech selects only
experts within a particular scientific field based on:

1. Field of research
2. Publication history - a record of scientific accomplishment documented by contribution to scientific
literature or other evidence of scientific accomplishment and
3. Relevance, quality and impact of past publications

After successfully meeting these criteria, Professor Ortiz-Rodriguez has been invited to participate in
the book project under the working title "Artificial Neural Networks", ISBN 978-953-307-243-2.
Sincerely yours,
Aleksandar Lazinica, CEO

T/F +385 51 315 152
E info@intechweb.org
www.intechweb.org

TO:
Jose Manuel Ortiz-Rodriguez

Mexico

DATE: Jul 06, 2010

INTECH NOTIFICATION OF ACCEPTANCE

Dear Prof. Jose Manuel Ortiz-Rodriguez,

On behalf of the Editorial Board it is my pleasure to inform you that the manuscript entitled "Evolutionary
Artificial Neural Networks in Neutron Spectrometry", has been accepted to be published in the book "Artificial
Neural Networks", ISBN 978-953-7619-X-X.

We firmly believe that your contribution will be of great importance for the scientific community.

Sincerely yours,
Aleksandar Lazinica, President and CEO

INTECH International Offices:

OFFICE RIJEKA:
Janeza Trdine 9
51000 RIJEKA
Croatia

OFFICE VIENNA:
Zieglergasse 14
A-1070 Vienna
Austria

DISTRIBUTION CENTRE:
Phoenix Consulting Services Ltd.
3/240 Vivek Khand, Gomti Nagar.
Lucknow-226010, India

0
Evolutionary Artificial Neural Networks in Neutron
Spectrometry
Jose Manuel Ortiz-Rodrgueza,c , Ma. del Rosario Martnez-Blancob and
Hector Rene Vega-Carrillob

U. Academicas: a Ing. Electrica, b Est. Nucleares, Universidad Autonoma de Zacatecas.

c Depto. de Electrotecnia y Electr
onica, Escuela Politecnica Superior, Cordoba Espana.

Mexico

1. Introduction
1.1 Artificial neural networks

ANN are highly simplified models of the working of the brain (Graupe, 2007; Kasabov, 1998).
An ANN is a biologically inspired computational model which consists of a large number
of simple processing elements called neurons, units, cells, or nodes which are interconnected
and operate in parallel (Galushkin, 2007; Lakhmi & Fanelli, 2000). Each neuron is connected
to other neurons by means of directed communication links, which constitute the neuronal
structure, each with an associated weight (Dreyfus, 2005). The weights represent information
being used by the net to solve a problem. Figure 1 shows an abbreviated notation for an
individual artificial neuron, which is used in schemes of multiple neurons (Beale et al., 1992).
Here the input p, a vector of R input elements, is represented by the solid dark vertical bar at
the left. The dimensions of p are shown below the symbol p in the figure as Rx1. These inputs
post multiply the single-row, R column matrix W. A constant 1 enters the neuron as an input
and is multiplied by a bias b. The net input to the transfer function f is n, the sum of the bias b
and the product Wp. This sum is passed to the transfer function f to get the neurons output
a.

Fig. 1. Abbreviated notation for an individual artificial neuron

Although a single neuron can perform certain simple information-processing functions, a single node is insufficient for many practical problems, and networks with a large number of
nodes are frequently used. A single layer of neurons having different transfer functions can
be created simply by putting the neuron shown earlier in parallel (Kishan et al., 2000). All the
neurons would have the same inputs, and each neuron would create the outputs, however,
a layer of neurons is not constrained to have the number of its inputs equal to the number
of its neurons, and it is common for the number of inputs to a layer to be different from the
number of neurons. To describe networks having multiple layers, it needs to make a distinction between weight matrices that are connected to inputs and weight matrices that are
connected between layers. It also needs to identify the source and destination for the weight
matrices. Weight matrices connected to inputs are called input weights (IW), whereas weight
matrices coming from layer outputs are called layer weights (LW). Further, superscripts are
used to identify the source (second index) and the destination (first index) for the various
weights and other elements of the network.
Figure 2 shows an abbreviated notation of a single layer of neurons. As can be seen from this
figure, the weight matrix connected to the input vector p is labeled as an input weight matrix
(IW1,1 ). The input vector elements enter the network through the weight matrix W. The row
indices on the elements of matrix W indicate the destination neuron of the weight, and the
column indices indicate which source is the input for that weight. Thus, the indices in w1,2
say that the strength of the signal from the second input element to the first (and only) neuron
is w1,2 . Elements of layer 1, such as its bias, net input, and output have a superscript 1 to say
that they are associated with the first layer.

Fig. 2. Abbreviated notation of a single layer of neurons

Multiple-layer networks are quite powerful and can solve more complicated problems than
can single-layer nets. A multilayer neural network consists of a combination of neurons or
nodes and synaptic connections, which are capable of passing data trough multiple layers
(Fausett, 1993). Each layer has a weight matrix W, a bias vector b, and an output vector a.
The layers of a multilayer network play different roles, i.e., the x-y-z neural network structure
refers to number of neurons in the input, hidden and output layers respectively. Input layers
receive input signal or values from an external source, output layer transmit the result of the
neural network processing and hidden layer(s) make up the internal layer(s) between input
and output node layers (Haykin, 1999). To distinguish between the weight matrices, output

vectors, etc., as mentioned previously, the number of the layer is appended as a superscript to
the variable of interest.
Figure 3 shows a three-layer network using abbreviated notation. From this figure can be seen
that the network has R1 inputs, S1 neurons in the first layer, S2 neurons in the second layer,
etc. A constant input 1 is fed to the bias for each neuron. The outputs of each intermediate
layer are the inputs to the following layer. Thus layer 2 can be analyzed as a one-layer network
with S1 inputs, S2 neurons, and an S2 xS1 weight matrix W2 . The input to layer 2 is a1 ; the
output is a2 . Now that all the vectors and matrices of layer 2 have been identified, it can be
treated as a single-layer network on its own. This approach can be taken with any layer of the
network.

Fig. 3. Three-layer neural network using abbreviated notation

The arrangement of neurons into layers and the connection patterns within and between layers is called the net architecture (Jain et al., 1996; Zupan, 1994). According to the absence or
presence of feedback connections in a network, two types of architectures are distinguished:
Feedforward architecture. There are no connections back from the output to the input
neurons; the network does not keep a memory of its previous output values and the
activation states of its neurons; the perceptron-like networks are feedforward types.
Feedback architecture. There are connections from output to input neurons; such a
network keeps a memory of its previous states, and the next state depends not only on
the input signals but on the previous states of the network; the Hopfield network is of
this type.
The central idea of neural networks, where w and b are both adjustable parameters of the neuron, is that such parameters can be adjusted by means of learning or training, so that the network exhibits some desired or interesting behavior (Graupe, 2007; Kishan et al., 2000; Lakhmi
& Fanelli, 2000; Haykin, 1999; Kasabov, 1998; Fausett, 1993). Learning is not an individual
ability of a single neuron, it is a collective process of the whole neural network and a result
of a training procedure. Training is the algorithmic procedure whereby the parameters of the
neurons of the network are estimated, in order for the neural network to fulfill, as accurately
as possible, the task it has been assigned.
As shows figure 4, an ANN is trained so that a set P of input vectors produces the desired,
or at least a consistent, set of target output vectors T, or the network learns about internal
characteristics and structures of data from a set P. The set P used for training a network is
called training set and the elements p of this set P are called training examples. The training
process is reflected in changing the connection weights of the network. The default performance function for feedforward networks is mean square error (mse), which is the average

squared error between the network outputs a and the target outputs T. During training, the
network weights should gradually converge to values such that each input vector p from the
data set training causes a desired output vector t produced by the network.
Learning occurs if after supplying a training example, a change in at least one synaptic weight
takes place. Input vectors and the corresponding target vectors are used to train a network
until it can approximate a function, associate input vectors with specific output vectors, or
classify input vectors in an appropriate way as defined by the experimenter.

Fig. 4. Training procedure of neural networks

The learning ability of a neural network is achieved through applying a learning or training
algorithm. Training algorithms are mainly classified into three groups:
Supervised. This training algorithm has been the most used mainly because the training examples comprise input vectors p and the desired target output vectors t. Training
is performed until the neural network learns to associate each input vector p to its
corresponding and desired output vector t.
Unsupervised. Only input vectors p are supplied; the neural network learns some
internal features of the whole set of all the input vectors presented to it.
Reinforcement learning. Sometimes called reward-penalty learning, is a combination
of the above two paradigms; it is based on presenting input vector p to a neural network
and looking at the output vector calculated by the network. If it is considered good,
then a reward is given to the network in the sense that the existing connection weights
are increased; otherwise the network is punished, the connection weights, being considered as not appropriately set, decrease. Thus reinforcement learning is learning
with a critic, as opposed to learning with a teacher.
Several different training algorithms for feedforward networks use the gradient of the performance function to determine how to adjust the weights to minimize performance. The gradient is determined using a technique called Back-Propagation (BP), which involves performing
computations backward through the network, which refine one of the principal components
of neural networks: the connection weights. The BP computation is derived using the chain
rule of calculus (Taylor, 1993). BP was created by generalizing the Widrow-Hoff learning rule
to multiple-layer networks and nonlinear differentiable transfer functions (Graupe, 2007; Kishan et al., 2000; Lakhmi & Fanelli, 2000; Haykin, 1999; Kasabov, 1998; Fausett, 1993).
Despite the apparent success of the BP learning algorithm, there are some aspects, which make
the algorithm not guaranteed to be universally useful.

One of the problems of BP is that it can get stuck in a local minimum. This is not too
bad if the local minimum turns out to be close to the global minimum, but there is no
guarantee that is the case.
Another problem associated with BP is that the place at which one starts on the error
surface (which is determined by the initial weight settings, which are often random)
determines whether or not a good or the best solution is found. When a solution is
found that performs well on the training set, the network might still perform badly on
the overall set of input, if the training set was not representative.
A last problem is the occurrence of interference. This occurs when a network is supposed to learn similar tasks at the same time. Apart from the fact that smaller networks
are unable to learn to many associations, they simply are full after a certain amount of
learned associations, there is also the danger of input patterns being so hard to separate,
that the network cant find a way to do it.
ANN has been well known for its effectiveness in representing nonlinear process system
(Apolloni et al., 2009). The power of neural computation comes from connecting neurons
in networks, and the way these nodes are connected, determines how computations proceed,
and constitutes an important early design decision by a neural network developer. However,
ANN can not solve all problems in the real world. One of the biggest drawbacks in the use
of neural networks nowadays is the problem of finding an appropriate structure for a given
task, mainly because it is very hard to know beforehand the size and the structure of a neural network one needs to solve a given problem (Ortiz-Rodrguez et al., 2006; Packianather &
Drake, 2004; Packianather et al., 2000; Peterson et al., 1995). An ideal structure is a structure
that independently of the starting weights of the net, always learns the task, i.e. makes almost
no error on the training set and generalizes well.
The problem with neural networks is that a number of parameter have to be set before any
training can begin. However, there are no clear rules how to set these parameters. Yet these
parameters determine the success of the training. Among the limitation of ANN, the followings should be given added emphasis:
1. Network architecture. There is a lack of fixed rule or systematic guideline for optimal
ANN architecture design. Since there is no a prior knowledge about the problem complexity, the network architecture was typically set arbitrarily. The network topology
was often determined by trial and error. This subjected the network to performance
uncertainties since the size of network influence the network performance: too small a
network cannot learn well, but too large may lead to overfitting. Thus, algorithms that
can find appropriate network architecture are needed. This includes the determination
of optimum number of neuron in each layer as well as number of hidden layer needed.
2. Training algorithm. The best training algorithm still cannot be singled out for general
neural network. Although BP algorithm has been widely used, it does not guarantee
the global optimal solution. The training may result in ANN model that is only accurate
in the same operating zones as in the training data set but inaccurate in others. Besides,
the selection of some parameters in BP training, such as learning rate and momentum,
also lacks of systematic guideline.
3. Training data. The quality and quantity of training data is an important issue for ANN
modeling. Usually, the success of ANN relies heavily on a large amount of data, but this
demand more computing time for training. In order to reduce the amount of data whilst

maintaining the model quality, the data used must be carefully selected to ensure that
they are sufficiently rich. This demands project understanding on the process involved.
Additionally, to eliminate noise and outliers, process data may require pre-processing
prior to application in neural network model development.
4. Training set. Since it is normally impossible to present a network with all possible
inputs, we only present it with part of it, the training set. This set has to be chosen in
such a way that the network also gives correct output for an input that was not in the
training set. If the network also responds well to inputs that were not in the training
set, it is said to generalize well. Often an ANN is trained with one set of patterns (the
training set) and tested with another (the test set). If the training set was not a good
representation of all possible inputs, the network probably will not perform too well on
inputs that are not in the training set. Generalization is quite similar to interpolation in
mathematics.
5. Process relationship. Being black-box method for modelling, ANN is criticized for
unable to explain and analysis the relationship between inputs and outputs. This may
cause difficulties in interpreting results from the network.
All of these limitations have motivated researchers to generate ideas of merging or hybridizing ANN with other approaches in the search for better performance. Some of the available schemes include expert systems, statistical methods (Ortiz-Rodrguez et al., 2006; Packianather & Drake, 2004), fuzzy logic (Chennakesava, 2008), wavelet transform and as well
as Neuro Evolutionary (NE) Approaches (Floreano & Mattiussi, 2008; Melin & Castillo, 2005;
Leardi, 2003; Yao, 1993). In this work, the use of NE is considered.
1.2 Neuroevolution

NE leverages the strengths of two biologically inspired areas of Articial Intelligence (AI)
(Rasskin-Gutman, 2009; Luger, 2005; Coppin, 2004): Artificial Neural Networks (ANN) and
Evolutionary Algorithms (EA) (Gen et al., 2009; Munakata, 2008; Rothlauf, 2006; Whiteson &
Stone, 2006). EA are stochastic and adaptive population-based search methods based on the
principles of natural evolution. They involve a population of individuals represented in a
genotypic form (chromosomes/genotypes), each of which is a potential solution to the problem. Each individual has a fitness score associated with it, and individuals with better fitness
scores are better solutions. Between one generation and the next, individuals are selected from
which to create offspring by applying mutation and crossover operators. Generally selection
is biased towards fitter individuals, and unpromising areas of the search space are abandoned
with the loss of poorer performing individuals from the population over time. EA encompass
Genetic Algorithms (GA), Evolutionary Programming (EP) and Evolution Strategies (ES) (Affenzeller et al., 2009; Haupt & Haupt, 2004; Mitchell, 1998; Periaux & Winter, 1995; Goldberg,
1989).
Although EA and ANN have in common that they are general search strategies, they vary
in their range. EA perform a more global search than ANN with BP. Figure 5, ilustrates the
convergence of the strategies. BP takes more time to reach the neighbourhood of an optimal
solution, but then reaches it more precisely. On the other hand, EA investigate the entire
search space. Hence, they reach faster the region of optimal solutions, but have difficulties
to localize the exact point. This happens, because for the final fine-tunning of the solution
relies almost enterely on mutation. Combining both strategies seems to be the best thing to
do (Floreano & Mattiussi, 2008; Melin & Castillo, 2005; Leardi, 2003; Yao, 1993), and the NE
approach outperforms EA as well as ANN in finding a satisfying solution.

Fig. 5. Convergence of ANN and EA

Because GAs sample many points in the search space simultaneously, they are less susceptible to local minima than single solution methods (BP), and are capable of rapidly locating
high payoff regions of high dimensional search spaces. Figure 6, shows a hypothetical fitness
landscape to illustrate how a GA operates. The fitness of each individual in the population is
represented by its position on the landscape. In a single solution method, such as BP training,
if the initial search point (the yellow circle) happens to fall in the neighborhood of a local maxima, the algorithm can become trapped because it has only local information with which to
make a next guess and improve the solution. Therefore, it will climb the gradient towards the
local maxima. In a GA, although some individuals (the red circles) may reside near local maxima, it is less likely to get trapped because the population provides global information about
the landscape. There is a better chance that some individual will be near the global maxima,
and the genetic operators allow the GA to move the population in large jumps to focus the
search in the most fruitful regions of the landscape.

Fig. 6. Fitness landscape which ilustrates how a GA operates

As is showed in figure 7, the basic idea of NE is to search the space of neural network policies
directly by using a GA. From this figure can be seen that each chromosome is transformed
into a neural network phenotype and evaluated on the task. The agent receives input from
the environment (observation) and propagates it through its neural network to compute an
output signal (action) that affects the environment. At the end of the evaluation, the network
is assigned a fitness according to its performance. The networks that perform well on the task
are mated to generate new networks.

Fig. 7. Neuro Evolution approach

GA were introduced by John Holland in 1975 (Goldberg, 1989), and are a class of stochastic
search procedures founded on the principles of natural selection. Unlike conventional search
methods that iteratively improve a single solution, a GA maintains a set or population of candidate solutions that sample the search space at multiple points. These solutions are encoded
as strings called chromosomes that represent the genotype of the solution. The chromosomes
are usually composed of a fixed number of genes that can take on some set of values called
alleles.
1.2.1 Genetic Algorithms

GA belongs to a class of population-based stochastic search algorithm that are inspired from
principles of natural evolution known as EA. Similar to other EA algorithms, GA is based on
the principle of survival of fittest, as in the natural phenomena of genetic inheritance and
Darwinian strife for survival. In other words, GA operates on a population of individuals
which represent potential solutions to a given problem. Mimicking the biological principles
in nature, a single individual of a population usually is affected by other individuals as well
as the environment. Normally, the better an individual performs under these competitive
conditions the greater is the change for the individual to survive and reproduce. This in turn
inherits the good parental genetic information. Hence, after several generations, the bad individual will be eliminated and better individuals are produced.
The most important terms used in the GA, analogous to the terms used to explain the evolutionary processes, are:
Gene. A basic unit, which controls a property of an individual.

Chromosome. A string of genes; it is used to represent an individual, or a possible

solution of a problem in the solution space.
Population. A collection of individuals.

Operation of crossover or mating. Substrings of different individuals are taken and

new strings (offsprings) are produced.
Mutation. Random change of a gene in a chromosome.

Fitness or goodness function. A criterion which evaluates each individual.

Selection. A procedure for choosing a part of the population that will continue the
process of searching for the best solution, while the other part of the population dies.

Although the evolutionary principle of GA is similar with other EA varieties, its implementation is different. Initially, the evolutionary process of GA starts with the creation of the blind
random population defined by the problem statement. This is followed by the series of activity including solutions encoding, fitness evaluation, selection, genetic operator alteration and
replacement which are iteratively executed until the stopping criterion is satisfied. Figure 8
outlines a typical evolutionary structure of GA.

Fig. 8. Evolutionary structure of a GA

As a population-based search algorithm, information exchange among the individual is particularly important for GA. Such mechanisms are achieved by genetic operators, the more
common ones are reproduction, crossover and mutation. Following a process analogous to
natural evolution, each genotype is transformed into its phenotype and evaluated on a given
problem to assess its fitness. Those genotypes with high fitness are then mated using crossover
and mutation at low levels to produce new solutions or offspring. Figure 9 illustrates how
crossover and mutation work. Crossover produces two offspring from two parents by exchanging chromosomal substrings on either side of a random crossover point, each offspring
is a concatenation of contiguous gene segments from both parents. When an offspring is mutated, one of its alleles is randomly changed to a new value. By mating only the most fit
individuals, the hope is that the favorable traits of both parents will be transmitted to the
offspring resulting in a higher scoring individual, and eventually leading to a solution.

Fig. 9. Crosover and mutation genetic operators

In general, GA is applicable to a wide range of problem in learning and optimization. They can
deal with complex problems which are multimodal and discontinuous. GA has two prominent features that are different from other search algorithms. First, it is population-based.
Second, there is information exchange among individuals in a population. Primarily, GA was
designed to optimally solve sequential decision processes more than to perform function optimization but over the years, it has been used widely in both learning and optimization. For
these reasons, GAs are well suited for searching the space of neural networks. Instead of training a network by performing gradient descent on an error surface, the GA samples the space
of networks and recombines those that perform best on the task in question.
Today neural networks can be trained to solve problems that are difficult for conventional
computers or human beings, and have been trained to perform complex functions in various
fields, including pattern recognition, identification, classification, speech, vision, and control
systems. Recently, the use of ANN technology has been applied with success in the research
area of nuclear sciences, mainly in the neutron spectrometry and dosimetry domains.
1.3 Artificial neural networks and neutron spectrometry

Nowadays, neutrons are widely used in many fields of both research and technology (Wielunski et al., 2008b). Reliable determination of neutron doses is still an issue in the field of radiation protection (Lacoste et al., 2007). In recent years, the characterization of ionizing radiation
fields in workplaces is one of the challenging activities over the world (Mazrou et al., 2008).
The workers subject to these radiations especially those who are submitted to neutron risk
have to be well monitored and protected according to relevant national regulations which are
more and more restrictive. As a result, there is an increasing demand in the field of radiation protection to quantify these various neutron fields and to determine the radiation doses
involved (Mazrou et al., 2008). The dosimetry of neutron radiation is one of the most complicated tasks in radiation protection (Wielunski et al., 2008a), mainly because is a complex
technique (Bedogni et al., 2007), and highly neutron energy dependent, and a precise knowledge on neutron spectrometry is highly essential for all dosimetry-related studies as well as
many nuclear physics experiments. In consequence, it becomes necessary to develop additional measuring techniques to enhance the actual workers monitoring systems.
The term radiation spectrometry can be used to describe measurements of the intensity of a
radiation field with respect to energy, wavelength, momentum, mass, angle of incidence or

any other related quantity (Vega-Carrillo et al., 2010; 2009a; Thomas, 2004). The distribution
of the intensity with one of these parameters is commonly referred to as the spectrum, i.e.,
the measurement of neutron energy spectra is the variation of the intensity of these radiations
with energy (Vega-Carrillo et al., 2009b).
One of the suitable approaches to improve the knowledge on neutron radiation fields to which
individuals are exposed during their work, is based on spectrometric measurements (Thomas,
2004; Brooks & Klein, 2002; McDonald et al., 2002). The measured yield of a radionuclide is the
convolution of the neutron energy distribution with the response function of the spectrometer
summed over the interaction energy range. At any given angle, the measured activity of a
particular type is related to the neutron spectrum by equation 1, which represents the discrete
form of the Fredholms integral equation of the first kind without any unique solution (VegaCarrillo et al., 2006).
Cj =

Ri,j i > j = 1, 2, ..., m

(1)

i =1

where Cj is jth detectors count rate; Ri,j is the jth detectors response to neutrons at the ith
energy interval; i is the neutron fluence within the ith energy interval and m is the number
of spheres utilized.
Although there is a wide range of different devices used for neutron spectrometry, the majority of the instruments can be grouped together into a small number of broad categories,
each one based on a common underlying technique (Thomas, 2004; Matzke, 2003; McDonald
et al., 2002). Among the many available neutron spectrometry techniques, the multisphere or
Bonner sphere spectrometer (BSS) system is the most used for radiation protection purposes
(Lacoste et al., 2004; El Messaoudi et al., 2004; Vylet, 2002), due to advantageous characteristics as wide energy range (from thermal to GeV neutrons), large variety of active or passive
thermal sensors allowing adapting the sensitivity to the specific workplace, good photon discrimination and simple signal management. Disadvantages are the poor energy resolution,
which does not allow appreciating fine structures as narrow peaks, the weight, and the need
to sequentially irradiate the spheres, requiring, in general, long exposure periods (Bedogni
et al., 2007).
The BSS consists of a thermal neutron sensor such as 6 LiI ( Eu), which is placed at the centre
of a number of moderating spheres of different diameter. With the BSS, neutron spectrum can
be obtained, however, the derivation of the spectral information is not simple. The unknown
neutron spectrum is not given directly as a result of the measurement. The BSS response matrix, the count rates and the neutron spectrum are related trough the equation 1. The most
delicate part in the neutron spectrometry based on the BSS, is the unfolding process. The
unfolding spectra of the neutrons measured consist on establishing the rate of energy distribution of fluency ( E), known as the response matrix, Ri,j , and the group of carried out
measures, Cj . Because the number of incognits overcormes to the number of equations, this
is an ill-conditioned system and has an infinite number of solutions. The procedure of selecting the solution that has meaning for the problem type, is part of the unfolding process
(Vega-Carrillo et al., 2005; 2006). The spectral information needs to be unfolded from the BSS
system detector responses by using a suitable computational code, most of them are based in
some of these methods: least square, iterative (Bedogni et al., 2007; Miller, 1993), bayesian and
maximum entropy (Reginatto & Zimbal, 2008), and Monte-Carlo (Vega-Carrillo et al., 2007a).
The current interest in the neutron spectrometry problem, has stimulated the development
of diverse unfolding procedures which try to obtain a better energy resolution through the

reconstruction of the spectrum. During the past decades have been carried out intents to
develop new neutron spectra unfolding codes like BUNKIUT (Miller, 1993), FRUIT (Bedogni
et al., 2007), MAXED, UMG (Roberts, 2007), etc., to attain improved energy resolution through
spectrum unfolding. However, these methods still present the serious drawback of requiring
a very expert user for their operation and the necessity to provide an initial guess spectrum
for the deconvolution of the spectrum. To overcome these drawbacks, alternative approaches
have been studied and proposed, to make an efficient neutron dosimetry, and several unfolding procedures combined with various types of experimental methods have been reported
such as Genetic Algorithms (GA) (Mukherjee, 2004; Freeman et al., 1999), and Artificial Neural Networks (ANN) (Vega-Carrillo et al., 2010; 2009a; 2007b; 2006; 2005).
Many of the previous studies in neutron spectrometry and dosimetry by using the ANN
approach have found serious drawbacks in the ANN design process itself, mainly in the
proper determination of the structural and learning parameters of the networks being designed (Ortiz-Rodrguez et al., 2006). These parameters are significant contributing factors to
the ANN performance, however, the optimal selection of these parameters follows in practical
use no rules, and their value is at most arguable, mainly because they are generally heuristically chosen by using the trial and error technique, which produces poor artificial neural
networks with low generalization capacity and poor performance. For the anterior, the nuclear research community needs approaches that implement ANN models faster than what is
currently available. In consequence, more research has been suggested in order to overcome
these drawbacks (Vega-Carrillo et al., 2010; Bedogni et al., 2007; Vega-Carrillo et al., 2007b;
2006).
At present, one promising technique to design the structural and learning parameters of ANN
is by introducing adaptation of network training using EA. EA seems to be a proper alternative
to solve the ANN optimization problem and can be used to assist in the ANN design and
training (Ortiz-Rodrguez et al., 2010a; 2009b; 2008). However, as a novel approach in the
nuclear sciences area, the lack of information and tools for the analysis of the results obtained
whit these new technologies, makes difficult the work in this research area.
The aim of the present work is focused in analyzing the intersection of ANN and GA, analyzing like it is that GA can be used to help in the design processes and training of ANN, i.e.,
in the optimum selection of the structural and learning parameters of ANN, improving its
generalization capacity, in such a way that the neural network designed is able to unfold in an
efficient way neutron spectra, starting only from the count rates obtained with a BSS system.
Because the novelty of EANN technology in neutron spectrometry, lack of tools for the
analysis is observed. For this reason, an unfolding code based on EANN technology, devoted
to the operational workplace neutron monitoring, would be of great help to the radiation
protection community. With this purpose, in this work a new computer tool based on EANN
technology called Neutron Spectrometry and Dosimetry based on Evolutionary Artificial
Neural Networks (NSDEann), was developed in a customized front end user capable to
unfold neutron spectra and to simultaneously calculate 13 equivalent doses, by using only
as input data the count rates coming of a BBS system, in just a few seconds if compared with
the time spent with the classical techniques, not being needed a priori information about the
spectra being calculated.

2. Materials and Methods

In this work, EANN technology was used for the modeling and optimization of ANNs capable
to solve the neutron spectra unfolding problem, starting from the count rates coming from a
BSS with a 6 LiI ( Eu) thermal neutron detector, 7 polyethylene spheres of 0, 2, 3, 5, 8, 10, and, 12
inches of diameter respectively, and the UTA4 response matrix expressed for 31 energy bins
(Vega-Carrillo et al., 2009a). In this study, two neutron spectra were produced using a 239 PuBe
neutron source (Vega-Carrillo et al., 2009b). A neutron spectrum was produced by the 239 PuBe
neutron source located at 100 cm of distance, in an open space at 200 cm above floor level,
and was measured using the BSS previously described. Then, the same neutron source was
inserted in a cilindrical container with water, and the neutron spectrum was measured again.
The count rates measured in both cases with the BSS, were utilized to unfold the neutron
spectra and to calculate the dosimetric features using four EANN designed for each case,
varying some GA and ANN parameters during training stage. To analyze the performance
and generalization capability of the evolved networks, a new neutron spectra unfolding code
denominated Neutron Spectrometry and Dosimetry based on Evolutionary Artificial Neural
Networks (NSDEann), was used. Also, the spectra were unfolded with the BUNKIUT code
and UTA4 response matrix (Vega-Carrillo et al., 2009a), and compared with that obtained with
NSDEann code using a comumputer tool known as Neutron Spectrometry and Dosimetry
Tool Box (NSDTB) (Ortiz-Rodrguez et al., 2010b; 2009a;b).
The general idea of combining EA and ANN is ilustrated in figure 10. The EANN approach is
based on a fundamental cyclic process which consists of:
1. Creating an initial population of genotypes (genetic representations of the ANN).
2. Building neural networks (phenotypes) based on the genotypes.
3. Training and testing the neural networks to determine how fit they are.
4. Comparing the fitness of the networks and keeping the best.
5. Selecting those networks in the population which are better, discarding those which
arent.
6. Refilling the population back to the defined size.
7. Pairing up the genotypes of the neural networks.
8. Mating the genotypes by exchanging genes (features) of the networks.
9. Mutating the genotypes in some random fashion; Then returning back to step (2) and
continuing this process until some stopping criteria is reached or manually stops the
process.
Through the process described previously, the better networks survive and their features carry
forward into future generations and are combined with others to find better and better networks for the problem considered. This genetic search capability is much more effective than
random searching, as the genetic process of recombining features vastly improves the speed
of identifying highly fit networks.

Fig. 10. General approach of Evolutionary ANN

To train the evolved networks, a data set of 187 neutron spectra, compiled by the International

Atomic Energy Agency (IAEA) (IAEA, 2001) was used (Vega-Carrillo et al., 2006; 2005; Iniguez
de la Torre & Vega Carrillo, 2002). The ANN genetically evolved were designed by meas of
the NeuroGenetic Optimizer (NGO) software (BiocompSystems, 2010). In the use of NGO,
there are 10 working steps as follows:
1. Initial population. The first step is started from building new population consisted of
chromosomes derived from random sampling by having total chromosomes as population size.
2. Decode. This step is involved with decoding heredity from chromosomes derived from
population in the 1st step by transforming them in hidden neuron in the hidden layer
of ANNs.
3. Train Back-propagation Neural Network. In this step, network for learning would
be derived from the past data by assigning numbers of hidden neuron as equal as the
value from decoding. Before data could be applied in learning, data must be made into
normal design. Afterwards, data must be divided into 2 set, one in training, another
in testing. Learning process is the reverse learning process with variance that could be
reversed back to adjust weight so that variance may be reduced.
4. Fitness evaluation. In this step, calculation is done to find variance of network which
considered proper fitness value of heredity by using the Root-Mean-Square Error of
Prediction (RMSEP) as objective function. RMSEP is the square root of the sum of the
squared differences between the observed and predicted values for all observations in
the test set divided by the number of such observations, to estimate the prediction error,
as ishowed in the following equation:

RMSE =

1
n (y ybi )2
n 1 i =1 i

(2)

where n is the number of trials, yi is the measured values of each response and ybi is the
neural model output.

5. Stop criteria. This is the step to check when the design stops working by setting up
conditions for numbers of result compilation. If answers in each round are still stable,
it should stop working. For examples, stop at 10 rounds of compilation or receiving
duplicate answers 5 times in the row. If the stop conditions are real, operation may stop
at step 10. On the contrary, for unreal stop conditions, proceed further with step 6.
6. Selection. In this step, two chromosomes with the least fitness values are selected from
population to be the breeders.

7. Crossover. This is the step for crossing species by exchanging genes among selected
breeders in Step 6 for offspring.
8. Mutation. After deriving at offspring with 2 chromosomes after crossover, mutation
would be done by randomly selected position of gene with the possibility of mutation.
Then, after randomly selected position, value of gene would have opposite value.
9. Replacement. The offspring with fitness value would take place of the suitable chromosome in order to derive at new group of population before going back to 2nd step.
10. Stop. After the stop conditions had been verified, the operation of model would stop
with only the best network design had been collected for being used for measuring
effectiveness of network.
The following steps were carried out for developing four EANN models with NGO:
1. Identify data by dividing into 2 sets. Input layer data set, comprised for the count
rates coming from polyethylene spheres of BSS and, output layer data set, comprised
for neutron spectra unfolded and equivalent doses.
2. Specifying neuron in the structure of ANNs. Input layer has 7 neurons, corresponding
to spheres of BSS, hidden layer(s) use GAs to specify neurons and, output layer has 44
neurons, first 31 are for the neutron spectra unfolded and the remaining 13 for different
equivalent doses (Vega-Carrillo et al., 2010).
3. Input data classification. A data set with 187 neutron spectra was used. Training and
testing data sets were created by dividing the whole data set, by using a random procedure, into 80% for training and 20% for testing.
4. Neural Parameters. In this work, ANNs with supervised learning through BP were
chosen, the net architecture was optimized with the application of GAs to find suitable
networks. Sigmoid function was employed as transfer function for hidden layer(s) and
linear for output layer. The optimizer of the network structure, used all inputs, and
searched the neural architecture.
5. Genetic Parameters. In this study, as is showed in table 1, in four experimental cases,
two configuration parameters of GA were varied by using the trial and error technique:
the number of generations run (GR) and population size (PS). However, a major difficulty encountered when using GAs is the parameter setting (Pongcharoen et al., 2007).

There exist many forms and variations of GAs and the best choice is problem dependant on the proper selection of the genetic operators. A GA can show good or weak
results even when applied on the same problem. More research is needed in order to
overcome the drawbacks associated with the optimum selection of GA parameters.
NET GR PS
1
10
60
2
10
60
3
5
30
4
10
30
Table 1. GA configuration parameters

SEL
50%S
50%S
50%S
50%S

ROP
CS
CS
CS
CS

MAT
TST
TST
TST
TST

MUT
RET25%S
RET25%S
RET25%S
RET25%S

where GR are the Generations Run, PS is the Population Size, SEL is the Selection techinque, 50%S is the top 50% Surviving, ROP is the Refilling of the Population Technique,
CS is Cloning the Survivors technique, MAT is the Mating technique, TST is the TailSwap Technique, MUT is the Mutation technique and RET25%S is Random Exchange
Technique at a rate of 25%.
6. Training Parameters. As can be seen from table 2, by using the trial and eror technique,
in first three experiments, the minimum network training passes (MNTP) and the cutoff
for network training passes (CNTP) were selected between 200-250 respectively for each
network trained. In a similar way, in the last experiment, the same parameters were
selected between 50-55. The Limit on Hidden Neurons (LHN) was varied as well, as is
showed in table 2. In experiments one and three, each evolved network was trined five
and three times respectively and then the results were averaged. In all experiments, the
weight initialization was between 0-0.3, and the learning rate and momentum between
0.1-0.4 and 0.1-0.3 respectively. The performance of the EANN is highly dependant on
the proper selection of these parameters, however, from the literature reviewed, there is
not a methodological criteria for this selection, and more research is needed to overcome
this drawback.
NET MNTP CNTP
1
200
250
2
200
250
3
200
250
4
50
55
Table 2. EANN training configuration parameters

LHN
8
32
256
256

where MNTP are the minimum network training passes for each network, CNTP are
the cutoff for network training passes for each network, and LHN is the the Limit on
Hidden Neurons in the design process of the EANN methodology.
After optimum net tolopogies were determined, was observed that NGO presents several
inconveniences when applied in the neutron spectroscopy domain. First probelm is due the
way the compendium of neutron spectra of IAEA compilation was realized. These spectra
are normalized to one and are used to train the different EANNs. This does not represent
a problem in the training and testing stages of EANN design, however, after training was
done and the neural net is used to solve real experimental problems, this becomes a draback

because the spectra unfolded and doses calculated resulting are normalized to one and this
has not physical meaning. Because the anterior, in the solution of real experimental problems
a procedure to un normalize the spectra unfolded was needed. Another drawback when
NGO is used in the neutron spectrometry domain is that has not the capability to graph in a
proper way the spectra and doses calculated. The anterior suggested the necessity to design a
customized computer tool to overcome the drawbacks mentioned.
Because the novelty of EANN methodology applied in the neutron spectrometry research
and the lack of tools for the analysis of the spectrometric and dosimetric results obtained
with NGO, a customized computer code denominated Neutron Spectrometry and Dosimetry
based on Evolutionary Artificial Neural Networks (NSDEann), showed in figure 11, was
designed in a graphical user interface, under the LabVIEW programming environment, which
is easy, intuitive, friendly and quick in their use. This code is oriented to be used by the end
user in laboratory, experimental and/or research environments, and its aim is to overcome
the drawbacks associated with NGO when applied to solve the neutron spectra unfolding
problem.
The principle of operation of NSDEann is the following: after executing the main program, a
window, as showed in figure 11, will open. To unfold the spectra by using the NSDEann code,
the user should execute the following steps:
1. Select a file text with the rate counts measured with the BSS system through the BSS
counts rate tool. Prevously, the end user should be created a file with extension *.txt,
with the rate counts arranged in column form. In the case of several measurements, the
text file could contain several rate counts in colums tabulation separated. The tool Select BSS from file can be used to select one spectra from the file created just adjusting
the number of column selection.
2. Store the normalized rate counts in a file text with extension *.txt by means of the button
Save normalized counts. In this step, the program calculates a normalized value of
rate counts which has two purposes, normalize the BBS measurements to be used with
NGO and to un normalize the spectra and doses calculated.
3. Open the optimum EANN trained with NGO and make a prediction with this configuration, using the normalized rate counts calculated in step two.
4. Store the neutron spectra unfolded and doses calculated by NGO, which are normalized
to one, saving this information in a similar way than step one.
5. Open the file with the spectra and doses calculated with NGO by means of the tool
Evolved Spectra&Doses. This tool works similar to described in step one.
6. Click the button Plot Spectra&Doses to see the spectrometric and dosimetric information un normalized in the graph and numeric form in the middle and right areas of
the main window.
7. The numerical values of the spectra and doses can be stored in a file text using the
button Save Spectra&Doses
For each spectra, the procedure before described must be repeated.

Fig. 11. Main window of NSDEann unfolding code

As described previoulsy, four EANN architectures were designed with NGO for two experimental cases: the neutron spectra of a 239 PuBe neutron source in water and on air. To compare
the several spectra obtained in both experiments and to analyze the performance and generalization capability of the nets designed, the tool known as Neutron Spectrometry and
Dosimetry ToolBox (NSDTB), was used (Ortiz-Rodrguez et al., 2010b; 2009a). NSDTB code
has the capability to analyze spectrometric and dosimetric information obtained with several
neutron unfolding techniques as for example: iterative procedures and the approaches based
on ANN or EANN, as is the case of the present work.

3. Results
In this work, the EANN technology was analyzed and applied in the neutron spectrometry research area. The spectrometric and dosimetric feateures of a 239 PuBe neutron source measured
under two different experimental conditions were unfolded by using the NSDEann unfolding
code. The results obtained in this work reveals that the hybrid technology of EA-ANN applied in the neutron spectrometry and dosimetry problems, present some drawbacks in the
optimum selection of the GA and ANN training parameters. When these parameters were
varied before GA execution, in the four experimental cases considered, different network architectures were obtained, as is showed in table 3. As can be seen from this table, a total of
1650 evolutionary network architectures were designed, trained and tested by means of EA,
under four different experimental conditions. The accuracy on training and testing, is similar
in the four cases considered, being around 97%.

NET
NET-TOP
ATrS
MATtS
1
7-8-44
98.67%
91.89 %
2
7-19(1)-44
98.00%
94.59%
3
7-183(14)-123(39)44 100.00% 97.30%
4
7-66-44
97.33%
94.59%
Table 3. Evolved network topologies and performance

NT
600
600
150
300

TIME
06:04:49
01:54:30
04:42:35
01:16:34

FOG
9
10
4
10

where NET-TOP is the optimum evolved net topology, ATrS is the Accuracy on Training Set,
MATtS is the Maximum Accuracy on Test Set, NT are the total number of Networks Trained,
TIME is the time used by GA to train all net topologies and FOG is the best network Found
on Generation.
The accuracy on training and testing reveals that the four network architures learned well
the training and testing data sets, which lets infer that the performance and generalization
capability of all networks architectures is good, which was confirmed by using the NSDTB
code, as is showed in figure 12. In this figure can be seen the four neutron spectra unfolded
with the four network architectures showed in table 3, for a 239 PuBe neutron source measured
at 1m in water. Here, spectra-1 and doses-1 corresponds to the spectra unfolded with the
architecture of the NET 1, spectra-2 and doses-2 corresponds to the spectra unfolded with the
architecture of the NET 2, etc. From this figure can be seen that although the four EANN with
different architectures were used to unfold the neutron spectra of the source before mentioned,
by using in all cases the same rate counts measured with the BSS system, the results are similar.
By comparing the spectra calculated with the codes NSDEann and BUNKIUT, as is showed
in figure 13, can be seen that are alike. The spectra have a peak in the thermal region and
between 0.1 and 10 MeV (Vega-Carrillo et al., 2010).

Fig. 12. Nueutron spectra unfolded with the NSDEann code for a 239 PuBe at 1m in water

The NSDTB code realizes a chi square statistical test among the spectra analyzed, as is showed
at right side of the main window. As can be seen from table 4, Spectra-1 and Doses-1 are statistically compared whith the rest of spectra and doses of row 1, Spectra-2 and Doses-2 are
compared with spectra and doses of row 2, etc. This comparison reveals that there are statistical differences between most of the spectra unfolded with the diffrent network topologies,
mainly spectra 3 and 4. In the case of the equivalent doses, the test reveals that statisticaly
there are not differences between doses calculated with the four network architectures.
Chi sqare test compares two groups of data, as in the case of the energy bins that conform the
neutron spectra calculated with the several EANN architectures, however, when the energy
bins presents high variations in some of the values of the data being compared, as can be appreciated in figures 12 to 15, this test tends to fail. From the analysis realized in the present
work, has been observed that the chi square test presents some drawbacks, and a more accurate statistical test is needed.
NET
1
2
3
4

S1
0
9.634
7E+199
1E+201

S2
9.378
0
3E+199
3E+200

S3
53.355
22.043
0
8E+200

S4
58.761
45.063
1E+200
0

D1
0
1.373
1.972
1.624

D2
1.032
0
0.357
0.206

D3
1.366
0.359
0
1.007

D4
1.346
0.145
0.945
0

Table 4. Chi square test of NSDTB code for a 239 PuBe at 1m in water
Figure 13 shows a comparison of the four neutron spectra obtained with the NSDEann code
with respect to the neutron spectra calculated using the BUNKIUT code. Here, spectra-1 corresponds to the spectra calculated with the BUNKIUT code and the rest as was explained for
figure 12.

Fig. 13. BUNKIUT-NSDEann comparison of the spectra of a 239 PuBe at 1m in water

In the case of figure 13, the chi square test fails because the high variations of some of the
values of the energy bins that compose the several spectra compared, despite the similarity
betwen the shape and features of the spectra, as was explained for figure 12.
Figure 14 shows the neutron spectra unfolded for a 239 PuBe measured at 1m on air. As can
be seen from this figure, the spectra are very similar in shape and features, however, the chi
square test fails when compare the different spectra, as is showed in table 5.

Fig. 14. Nueutron spectra unfolded with the NSDEann code for a 239 PuBe at 1m on air
Table 5 shows the statistical test realized by the NSDTB code. As can be seen, despite the similarities of the spectra, the test fails. A more accurate stastical test for comparing the neutron
spectra obtained with several unfolding codes is needed.
NET
1
2
3
4

S1
0
14.892
7E+198
5E+201

S2
2E+200
0
1E+200
2E+201

S3
5E+201
108.52
0
4E+201

S4
6E+201
98. 579
5E+200
0

D1
0
0.27
0.162
1.888

D2
0.261
0
0.419
0.741

D3
0.153
0.413
0
1.893

D4
1.74
0.711
1.788
0

Table 5. Chi square test of NSDTB code for a 239 PuBe at 1m on air
Figure 15 shows a comparison of the four neutron spectra obtained with the NSDEann code
with respect to the neutron spectra calculated using the BUNKIUT code for a 239 PuBe measured at 1m on air. Here, spectra-1 corresponds to the spectra calculated with the BUNKIUT
code. As can be seen from this figure, the spectra are very similar in shape and features,
however, the chi square test fails when compare the different spectra.

Fig. 15. BUNKIUT-NSDEann comparison of the spectra of a 239 PuBe at 1m on air

4. Conclusions
In this work, the intersection of GA and ANN in the neutron spectrometry research by means
of ANN technology was analyzed. The most common and widely used methods of optimization in ANN are classical gradient methods, such as BP. These methods are characterized by
a very fast reaching the optimum, yet, they have a considerable disadvantage converging
the algorithm to the closest optimum. So never is known whether the result of the optimization is a local or a global optimum. On the other hand, the genetic search capability in ANN
design is much more effective than random searching approaches, as the genetic process of
recombining features vastly improves the speed of identifying highly fit networks. It also has
a potential advantage over just using personal experience in building neural networks, as new
and potentially better solutions may be found through this process than might be found using
the nearly unavoidable assumptions made by the user.
Contrary to the classical methods, the Evolutionary Algorithms cope with the global optimum
searching quite well, but they are not as precise as the classical methods. Another disadvantage of the EA is a big (sometimes very big) amount of compilations. A major difficulty encountered when using GAs is the parameter setting. There exist many forms and variations of
GAs and the best choice is problem dependant. Accordingly, a GA can show a good or weak
result even when applied on the same problem. Like other learning paradigms, the performance of GA is dependent on the parameter choice, on the problem representation and on the
fitness landscape.
The new ideas and concepts of EA and ANN bring new life into artificial intelligence research
applied in the nuclear ciences. However, new problems arise of combining EA and ANN,
such as the proper determination of the EA parameters or the need of computer tools to apply
this new technology.

Because the novelty of EANN technology in neutron spectrometry and the lack of tools for the
analysis, an unfolding code based on EANN technology, called Neutron Spectrometry and
Dosimetry based on Evolutionary Artificial Neural Networks (NSDEann), was developed in
a customized front end user capable to unfold neutron spectra and to simultaneously calculate
13 equivalent doses, by using only as input data the count rates coming of a BBS system, in
just a few seconds if compared with the time spent with the classical techniques, not being
needed a priori information about the spectra being calculated.
One disadvantage of the NSDEann code is the dependency with NGO software. Because the
anterior, a customized and independent code shold be of help in the neutron spectrometry
field where EANN technology is applied. At present, work is being done in this sense.
When the NSDTB code was used to compare the spectra unfolded with four evolutionary network architectures obtained with the NGO software and analyzed with the NSDEann unfolding code, was observed that the chi square statistical test failed. This is due of high variations
in some of the energy bins that compose the neutorn spectra. From the analysis realized in
the present work, has been observed that the chi square test presents some drawbacks, and a
more accurate stastical test for comparing the neutron spectra obtained with several unfolding
codes is needed.

5. References
Affenzeller, M., Winkler, S., Wagner, S. & Beham, A. (2009). Genetic algotithms and genetic
programming, modern concepts and practical applications, CRC Press.
Apolloni, B., Bassis, S. & Marinaro, M. (2009). New directions in neural networks, IOS Press.
Beale, M. H., Hagan, M. T. & Demuth, H. B. (1992). Neural networks toolbox, users guide, Mathworks. www.mathworks.com/help/pdf_doc/nnet/nnet.pdf.
Bedogni, R., Domingo, C., Esposito, A. & Fernandez, F. (2007). Fruit: an operational tool for
multisphere neutron spectrometry in workplaces, Nuclear Instruments and Methods in
Physics Research A 580: 13011309.
BiocompSystems (2010). Biocomp systems.
Brooks, F. D. & Klein, H. (2002). Neutron spectrometry, historical review and present status,
Nuclear Instruments and Methods in Physics Research A 476: 111.
Chennakesava, R. (2008). Fuzzy logic and neural networks, basic concepts and applications, New
Age International Publishers.
Coppin, B. (2004). Artificial intelligence iluminated, Jones and Barttlet Publishers.
Dreyfus, G. (2005). Neural networks, methodology and applications, Springer.
El Messaoudi, M., Chouak, A., Lferde, M. & Cherkaoui, R. (2004). Performance of three different unfolding procedures connected to bonner sphere data, Radiation Protection
Dosimetry 108(3): 247253.
Fausett, L. (1993). Fundamentals of neural networks, architectures, algorithms and applications,
Prentice Hall.
Floreano, D. & Mattiussi, C. (2008). Bio-inspired artificial intelligence, theories, methods and technologies, MIT Press.
Freeman, D. W., Edwards, D. R. & Bolon, A. E. (1999). Genetic algorithms, a new techniqe for
solving a neutron spectrum unfolding problem, Nuclear Instruments and Methods in
Physics Research A 425(3): 549576.
Galushkin, A. (2007). Neural networks theory, Springer.
Gen, M., Green, ., Katai, O., McKay, B., Namatame, A., Sarker, R. A. & Zhang, B. T. (2009).
Intelligent and evolutionary systems, Springer-Verlag.

Goldberg, D. (1989). Genetic Algorithms in search, optimization, and machine learning, Addison
Wesley.
Graupe, D. (2007). Principles of artificial neural networks, World Scientific.
Haupt, R. L. & Haupt, S. E. (2004). Practical genetic algorithms, Wiley.
Haykin, S. (1999). Neural networks: a comprensive foundation, Prentice Hall.
IAEA (2001). Compendium of neutron spectra and detector responses for radiation protection
purposes, Technical Report 403.

Iniguez
de la Torre, M. P. & Vega Carrillo, H. R. (2002). Catalogue to select the initial guess
spectrum during unfolding, Nuclear Instruments and Methods in Physics Research A
476(1): 270273.
Jain, A. K., Mao, J. & Mohiuddin, K. M. (1996). Artificial neural networks: a tutorial, IEEE:
Computer 29(3): 3144.
Kasabov, N. K. (1998). Foundations of neural networs, fuzzy systems, and knowledge engineering,
MIT Press.
Kishan, M., Chilukuri, K. & Sanjay, R. (2000). Elements of artificial neural networks, The MIT
Press.
Lacoste, V., Gressier, V., Pochat, J. L., Fernandez, F., Bakali, M. & Bouassoule, T. (2004). Characterization of bonner sphere systems at monoenergetic and thermal neutron fields,
Radiation Protection Dosimetry 110(1-4): 529532.
Lacoste, V., Reginatto, M., Asselineau, B. & Muller, H. (2007). Bonner sphere neutron spectrometry at nuclear workplaces in the framework of the evidos project, Radiation Protection Dosimetry 125(1-4): 304308.
Lakhmi, J. & Fanelli, A. M. (2000). Recent advances in artificial neural networks design and applications, CRC Press.
Leardi, R. (2003). Nature-inspired methods in chemometrics: genetic algorithms and artificial neural
networks, Elsevier.
Luger, F. G. (2005). Artificial Intelligence, structures and strategies for complex problem solving,
Addison-Wesley.
Matzke, M. (2003). Unfolding procedures, Radiation Protection Dosimetry 107(1-3): 155174.
Mazrou, H., Sidahmed, T., Idiri, Z., Lounis-Mokrani, Z., , Bedek, Z. & Allab, M. (2008). Characterization of the crna bonner sphere spectrometer based on 6lii scintillator exposed
to an 241ambe neutron source, Radiation Measurements 43: 10951099.
McDonald, J. C., Siebert, B. R. L. & Alberts, W. G. (2002). Neutron spectrometry for radiation protection purposes, Nuclear Instruments and Methods in Physics Research A 476(12): 347352.
Melin, P. & Castillo, O. (2005). Hybrid intelligent system for pattern recognition using soft computing, an evolutionary approach for neural networks and fuzzy systems, Springer.
Miller, S. (1993). Afitbunki: a modified iterative code to unfold neutron spectra from Bonner sphere
detector data, Masters thesis.
Mitchell, M. (1998). An introduction to genetic algorithms, MIT Press.
Mukherjee, B. (2004). Andi-03: a genetic algorithm tool for the analysis of activation detector data to unfold high-energy neutron spectra, Radiation Protection Dosimetry 110(14): 249254.
Munakata, T. (2008). Fundamentals of the new artificial intelligence, neural, evolutionary, fuzzy and
more, Springer.
Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R., Gallego, E. & Vega-Carrillo, H. R. (2008).
Artificial neural networks modeling evolved genetically, a new approach applied

in neutron spectrometry and dosimetry research areas, Proceedings of the Electronics,

Robotics and Automotive Mechanics Conference (CERMA08), IEEE Computer Society,
Cuernavaca, Mor., Mexico, pp. 387392.
Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R., Gallego, E. & Vega-Carrillo, H. R. (2009a).
A computational tool design for evolutionary artificial neural networks in neutron
spectrometry and dosimetry, Proceedings of the Electronics, Robotics and Automotive Mechanics Conference (CERMA09), IEEE Computer Society, Cuernavaca, Mor., Mexico,
pp. 113118.
Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R., Gallego, E. & Vega-Carrillo, H. R. (2009b).
Evolutive artificial neural networks for neutron spectra unfolding, American Nuclear
Society Transactions 101: 647648.
Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R., Gallego, E. & Vega-Carrillo, H. R. (2010a).
Neutron spectrometry and dosimetry based on a new approach called genetic artificial neural networks, 12th Congress of the International Radiation Protection Association
(IRPA12), IAEA Procceding Series STI/PUB/1460 1-9, International Atomic Energy
Agency, Vienna, pp. xxyy.
Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R., Gallego, E. & Vega-Carrillo, H. R. (2010b).
A neutron spectrometry and dosimetry computer tool based on ann, 12th Congress
of the International Radiation Protection Association (IRPA12), IAEA Procceding Series
STI/PUB/1460 1-9, International Atomic Energy Agency, Vienna, pp. xxyy.
Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R. & Vega-Carrillo, H. R. (2006). Robust design
of artificial neural networks applying the Taguchi methodology and DoE, Proceedings of the Electronics, Robotics and Automotive Mechanics Conference (CERMA06), IEEE
Computer Society, Cuernavaca, Mor., Mexico, pp. 131136.
Packianather, M. S. & Drake, P. R. (2004). Modelling neural network performance through
response surface methodology for classifying wood veneer defects, Proceedings of the
Institution of Mechanical Engineers, Part B 218(4): 459466.
Packianather, M. S., Drake, P. R. & H., R. (2000). Optimizing the parameters of multilayered
feedforward neural networks through taguchi design of experiments, Quality and
Reliability Engineering International 16: 461473.
Periaux, J. & Winter, G. (1995). Genetic algorithms in engineering and computer science, John Wiley
& Sons.
Peterson, G. E., St. Clair, D. C., Aylward, S. R. & E., B. W. (1995). Using taguchis method
of experimental design to control errors in layered perceptrons, IEEE Transactions on
Neural Networks 6(4): 949961.
Pongcharoen, P., Chainate, W. & Thapatsuwan, P. (2007). Exploration of genetic parameters
and operators through travelling salesman problem, Science Asia 33: 215222.
Rasskin-Gutman, D. (2009). Chess metaphors, artificial intelligence and the human brain, The MIT
Press.
Reginatto, M. & Zimbal, A. (2008). Bayesian and maximum entropy methods for fusion diagnostic measurements with compact neutron spectrometers, Review of Scientific Instruments 79(2): 398403.
Roberts, N. J. (2007). Investigation of combined unfolding of neutron spectra using the umg
unfolding codes, Radiation Protection Dosimetry 126(1-4): 398403.
Rothlauf, F. (2006). Representations for Genetic and Evolutionary Algorithms, Springer.
Taylor, J. G. (1993). Mathematical approaches to neural networks, North-Holland Mathematical
library.

Thomas, D. J. (2004). Neutron spectrometry for radiation protection, Radiation Protection

Dosimetry 110(1-4): 141149.
E., Gallego, E., Lorente,
Vega-Carrillo, H. R., Hernandez-Davila, V. M., Manzanares-Acuna,

A. & Iniguez,
M. P. (2007b). Artificial neural networks technology for neutron spectrometry and dosimetry, Radiation Protection Dosimetry 126(1-4): 408412.
E., Mercado Sanchez,
Vega-Carrillo, H. R., Hernandez-Davila, V. M., Manzanares-Acuna,

G. A., Gallego, E., Lorente, A., Perales-Munoz,

W. A. & Robles-Rodrguez, J. A.
(2005). Artificial neural networks in neutron dosimetry, Radiation Protection Dosimetry
118(3): 251259.
E., Mercado-Sanchez,
Vega-Carrillo, H. R., Hernandez-Davila, V. M., Manzanares-Acuna,

G. A., Iniguez
de la Torre, M. P., Barquero, R., Preciado-Flores, S., Mendez-Villafane,
R., Arteaga-Arteaga, T. & Ortiz-Rodrguez, J. M. (2006). Neutron spectrometry using
artificial neural networks, Radiation Measurements 41: 425431.
E., Ortiz-Rodrguez, J. M. & Arteaga-Arteaga, T.
Vega-Carrillo, H. R., Manzanares-Acuna,
(2007a). Neutron spectra re-binning and dose calculation using monte carlo methods,
Revista Mexicana de Fsica 53: 17.
Vega-Carrillo, H. R., Martnez-Blanco, M. R., Hernandez-Davila, V. M. & Ortiz-Rodrguez,
J. M. (2009a). Ann in spectroscopy and neutron dosimetry, Journal of Radioanalytical
and Nuclear Chemistry 281(3): 615618.
Vega-Carrillo, H. R., Ortiz-Rodrguez, J. M., Hernandez-Davila, V. M., Martnez-Blanco, M. R.,
Hernandez-Almaraz, B., Ortiz-Hernandez, A. & Mercado, G. A. (2009b). Different
spectra with the same neutron source, Revista Mexicana de Fsica S 56(1): 3539.
Vega-Carrillo, H. R., Ortiz-Rodrguez, J. M., Martnez-Blanco, M. R. & Hernandez-Davila,
V. M. (2010). Ann in spectroscopy and neutron dosimetry, American Institute of Physics
Proccedings x: xx. In press, accepted - July 2010.
Vylet, V. (2002). Response matrix of an extended bonner sphere system, Nuclear Instruments
and Methods in Physics Research A 476: 2630.
Whiteson, S. & Stone, P. (2006). Evolutionary function approximation for reinforcement learning, The Journal of Machine Learning Research 7: 877917.
Wielunski, M., Wahl, W., EL-Faramawy, N., Rhm, W., Luszik-Bhadra, M. & Roos, H. (2008a).
Development of a brazilian gamma-neutron dosimeter, Nuclear Instruments and Methods in Physics Research B 266(12-13): 31743177.
Wielunski, M., Wahl, W., EL-Faramawy, N., Rhm, W., Luszik-Bhadra, M. & Roos, H. (2008b).
Intercomparison exercise whith mev neutrons using various electronic personal
dosimeters, Radiation Measurements 43(2-6): 10631067.
Yao, X. (1993). Evolutionary artificial neural networks, International Journal of Neural Systems
4(3): 203222.
Zupan, J. (1994). Introduction to artificial neural network methods: what they are and how to
use them, Acta Chimica Slovenica 41(3): 327352.

REVISTA MEXICANA DE FISICA S 57 (1) 6971

FEBRERO 2011

Neutron spectrometry using artificial neural networks for a bonner sphere

spectrometer with a 3 He detector
J.M. Ortz-Rodrgueza,c, , M.R. Martnez-Blancob , H.R. Vega-Carrillob , E. Gallego Dazd , A. Lorente Fillold ,
R. Mendez Villafanee , J.M. Los Arcos Merinoe , and J.E. Guerrero Araquee
a
Unidad Academica de Ingeniera Electrica, Universidad Autonoma de Zacatecas,

e-mail: morvymm@yahoo.com.mx
b
Unidad Academica de Estudios Nucleares, Universidad Autonoma de Zacatecas,
Apartado Postal 336, Zacatecas, 98000, Mexico.
c
Depto. de Electrotecnia y Electronica Escuela Politecnica Superior,
Avda. Menendez Pidal s/n, Cordoba, Espana.
d
Universidad Politecnica de Madrid, Depto. de Ingeniera Nuclear, ETSI Industriales,
C. Jose Gutierrez Abascal, 2, 28006, Madrid, Espana.
e
CIEMAT, Laboratorio de Metrologa de Radiaciones Ionizantes,
Avda. Complutense, 22, 28040, Madrid, Espana.
Recibido el 10 de marzo de 2010; aceptado el 31 de agosto de 2010
Neutron spectra unfolding and dose equivalent calculation are complicated tasks in radiation protection, are highly dependent of the neutron
energy, and a precise knowledge on neutron spectrometry is essential for all dosimetry-related studies as well as many nuclear physics
experiments. In previous works have been reported neutron spectrometry and dosimetry results, by using the ANN technology as alternative
solution, starting from the count rates of a Bonner spheres system with a 6 LiI(Eu) thermal neutrons detector, 7 polyethylene spheres and the
UTA4 response matrix with 31 energy bins. In this work, an ANN was designed and optimized by using the RDANN methodology for the
Bonner spheres system used at CIEMAT Spain, which is composed of a 3 He neutron detector, 12 moderator spheres and a response matrix
for 72 energy bins. For the ANN design process a neutrons spectra catalogue compiled by the IAEA was used. From this compilation, the
neutrons spectra were converted from lethargy to energy spectra. Then, the resulting energy ?uence spectra were re-binned by using the
MCNP code to the corresponding energy bins of the 3 He response matrix before mentioned. With the response matrix and the re-binned
spectra the counts rate of the Bonner spheres system were calculated and the resulting re-binned neutrons spectra and calculated counts rate
were used as the ANN training data set.
Keywords: Neutron spectrometry; neutron dosimetry; neural networks.
La reconstruccion de espectros de neutrones y el calculo de los equivalentes de dosis dependen de la energa de los neutrones, y por ello es
esencial un conocimiento preciso de la espectrometra de los neutrones para realizar los estudios relacionados con la disimetra, as como para
la realizacion de muchos experimentos de la fsica nuclear. Estos no son problemas triviales y los investigadores han mencionado la necesidad
de desarrollar tecnicas de medicion adicionales para ampliar los actuales sistemas de monitoreo del personal laboral. En trabajos previos
se han reportado resultados relacionados con la espectrometra y la disimetra de neutrones, utilizando la tecnologa de redes neuronales
como solucion alternativa a partir de las tasas de conteo de un sistema de esferas Bonner con un detector de neutrones termicos de 6 LiI(Eu), 7
esferas de polietileno y la matriz de respuesta conocida como UTA4, expresada en 31 grupos de energa. En este trabajo, se diseno y optimizo
una red neuronal para el sistema de esferas Bonner del CIEMAT en Espana, utilizando la metodologa de diseno conocida como RDANN, el
cual esta compuesto de un detector de neutrones termicos de 3 He, 12 esferas moderadoras y una matriz de respuesta expresada en 72 grupos
de energa. Para el proceso de diseno de la red neuronal se utilizo un catalogo de espectros de neutrones compilado por la IAEA. A partir de
esta compilacion, los espectros de neutrones se convirtieron de letargia a energa. Posteriormente, e stos espectros se rebobinaron utilizando
el codigo MCNP de acuerdo con los grupos de energa de la matriz de respuesta del detector de 3 He. Con la matriz de respuesta y los
espectros rebobinados se calcularon las tasas de conteo del sistema de esferas Bonner y de esta forma, los espectros rebobinados resultantes
y las cuentas calculadas se utilizaron como el conjunto de datos de entrenamiento de la red neuronal.
Descriptores: Espectrometra de neutrones; dosimetra de neutrones; redes neuronales.
PACS: 29.30.Hs, 07.05.Mh, 29.40-n

Introduction

pational radiation exposure in neutron fields is mainly done

with multi-element systems where each element has a particular response to neutrons. With the neutron spectrum information and neutron fluence-to-dose conversion coefficients,
different dose quantities, like Hp or H (10), can be estimated.
With the Bonner spheres spectrometer (BSS), neutron
spectrum from thermal to several MeV can be obtained.

J.M. ORTIZ-RODRIGUEZ et al.

However, the weight, time consuming procedure, the need to

use an unfolding procedure and the low resolution spectrum
are some of the BSS drawbacks. The BSS response matrix,
the count rates and the neutron spectrum are related through
the Fredholm integral-diferential equation, whose discrete
version is.
Cj =

N
X

Ri,j i

j = 1, 2, . . . , m

(1)

i=1

where Cj is jth detectors count rate; Ri,j is the jth detectors response to neutrons at the ith energy interval; i is the
neutron fluence within the ith energy interval and m is the
number of spheres utilized.
Equation 1 is an ill-conditioned equations system with
an infinite number of solutions. To unfold the neutron spectrum, , several methods are used [1]. The ANN technology
is a useful alternative to solve this problem; however, several
drawbacks must be solved in order to simplify the use of these
procedures [2]. In the ANN design process, the choice of the
ANNs basic parameters often determines the success of the
training process. The selection of these parameters follows in
practical use no rules, and their value is at most arguable. The
ANN designers have to choose the architecture and determine
many of the parameters through the trial and error technique,
which produces ANN with poor performance and low generalization capability, spending often large amount of time.
In the aim to solve these drawbacks, a new approach known
as Robust Design of Artificial Neural Networks (RDANN)
methodology has been proposed [2]. In this work, an ANN
was designed and optimized by using the RDANN methodology for the BSS utilized by the Laboratorio de Metrologa
de Radiaciones Ionizantes (LMRI) of CIEMAT at Spain.

2. Materials and methods

CIEMATs BSS system is composed of a 3 He neutron detector, 12 moderator spheres and a response matrix. In this
work, the selected response matrix it is expressed for 72 energy bins. For the ANN design process a catalogue of neutron
spectra compiled by the International Atomic Energy Agency
(IAEA) was used [3]. These include spectra from isotopic
neutron sources, reference and operational neutron spectra
obtained from accelerators and nuclear reactors, and spectra
obtained from mathematical functions. From this compilation, the neutron spectra were converted from lethargy to energy spectra. Then, the resulting energy fluence spectra were
re-binned by using the MCNP code to the corresponding energy bins of the CIEMATs BSS system, and 13 equivalent
doses were calculated in the re-binning process [3].
Re-binned spectra were normalized to 1 neutron per second and the expected count rates in the Bonner sphere spectrometer were calculated using the CIEMAT response matrix. The count rates were utilized as inputs in the network,
while the respective neutron spectrum and equivalent doses
were utilized as the output, during the training and testing

stages. For the ANN design process, the net topology should
be determined, however, many factors affect the ANN performance. Over the past years, new proposals of ANN design procedures have been made, because the serious drawbacks due to the trial and error technique. Recently, a novel
approach known RDANN methodology, has been applied in
the design process of ANN. RDANN methodology is a powerful method based on parallel processes, where all the experiments are planned a priori, and the results are analyzed
after all experiments are completed. This is a systematic and
methodological approach of ANN design, based on statistical
methods, which maximize the ANN performance and their
generalization capacity.
In this work, a multilayer feed forward backpropagation
ANN with momentum and learning rate was designed. Using the RDANN methodology, the ANN was designed and
optimized to solve 14 independent problems: the neutron
spectra unfolding problem and the calculation of 13 different equivalent doses, starting only from the count rates of the
CIEMATs BSS system.

Results and discussion

The optimum network topology is 12:16:85, momentum:

0.001, learning rate: 0.1, training algorithm: trainscg, target MSE: 1e-4. The first layer (input) are the count rates
readed from the CIEMATs BSS, the second layer (hidden)
has 16 neurons and the thirth layer (output) has 85 neurons.
The first 72 outputs correspond to the neutron spectra, as is
showed in Fig. 1, and the remaining 13 are equivalent doses,
(Fig. 2). Training was carried using 200 spectra, randomly
selected from the whole data set, and the remaining were used
for testing the learning of the net. The neural net was trained
until the mean square error (MSE) was reduced to 104 , in a
time of 107 seconds average.

F IGURE 1. Neutron spectra unfolded with the optimized ANN.

Rev. Mex. Fs. S 57 (1) (2011) 6971

NEUTRON SPECTROMETRY USING ARTIFICIAL NEURAL NETWORKS FOR A BONNER SPHERE SPECTROMETER WITH. . .

From Figs. 1 and 2 can be seen that according the

RDANN methodology, the expected an obtained spectra and
doses pass the 2 test, which means that statistically both,
spectra and doses, are the same. From these figures can be
seen too, that the correlation test is close to the best value.
This means that this network topology has a high performance and generalization capability.

F IGURE 2. 13 equivalent doses calculated with the optimized

ANN.

Opposite to classical solution methods, in the ANN approach, the count rates comming from a BBS system are
the only input data, not being necessary initial guess spectrum. The neutron spectra is unfolded and simultaneously,
13 equivalent doses are calculated. The equivalent doses,
are not dependent of the spectrometric information, they are
calculated simulteneously, based only on the BSS counts.
Figs. 1 and 2 shown the results of one case obtained at the
testing stage of the designed ANN. Figure 1 shows the expected neutron spectra (red line) and the calculated by the
ANN (blue line). Similarly, Fig. 2, shows the 13 doses. The
RDANN methodology makes statistical tests as 2 and correlation in order to analyze the performance and generalization
capabilities of the network.

1. H.R. Vega-Carrillo, E. Manzanares Acuna, J.M. Ortiz

Rodrguez, and T. Arteaga Arteaga, Radiation Protection
Dosimetry 126 (2007) 408.
2. J.M. Ortiz-Rodrguez, M.R. Martnez-Blanco, and H.R. VegaCarrillo, Electronics, Robotics and Automotive Mechanics Conference 2 (2006) 131.

Conclusions

The use of ANN technology is a useful alternative to solve

the neutron spectrometry and dosimetry problems; however,
to obtain the best results, some drawbacks must be solved in
the ANN design process, such as the optimum ANN topology
selection. In this work, the ANN optimization methodology
known as RDANN, was used to design an ANN capable to
solve the neutron spectrometry and dosimetry problems for
the CIEMAT BSS system.
The neural net was trained and tested using a large set
of neutron spectra compiled by the IAEA. The success of
ANN technology in neutron spectrometry and dosimetry, using only the Bonner spectrometer count rates as input in the
trained network will overcome some of the the problems associated with the solution of such ill-conditioned problem.
The results here reported demonstrate that the use of this
technology has become in a useful tool.

Acknowledgements
This work was supported by Agencia Espanola de Cooperacion Internacional para el Desarrollo, AECID (Spain).

3. H.R. Vega-Carrillo, E. Manzanares-Acuna, V.M. HernandezDavila, R. Barquero, M.P. Iniguez, R. Mendez, F. Palacios, A.
Arteaga-Arteaga, T. and J.M. Ortiz-Rodriguez, Radiation Measurements 41 (2006) 425.

Rev. Mex. Fs. S 57 (1) (2011) 6971

REVISTA MEXICANA DE FISICA S 57 (1) 8992

FEBRERO 2011

Performance of artificial neural networks and genetical evolved artificial neural

networks unfolding techniques
J.M. Ortz-Rodrgueza,c, , M.R. Martnez-Blancob , H.R. Vega-Carrillob , E. Gallego Dazd , A. Lorente Fillold ,
R. Mendez Villafanee , J.M. Los Arcos Merinoe , J.E. Guerrero Araquee
Unidades Academicas: a Ingeniera Electrica,

e-mail: morvymm@yahoo.com.mx
b
Estudios Nucleares, Universidad Autonoma de Zacatecas,
Apartado Postal 336, Zacatecas, 98000, Mexico.
c
Depto. de Electrotecnia y Electronica Escuela Politecnica Superior,
Avda. Menendez Pidal s/n, Cordoba, Espana.
d
Universidad Politecnica de Madrid, Depto. de Ingeniera Nuclear, ETSI Industriales,
C. Jose Gutierrez Abascal, 2, Madrid, 28006, Espana.
e
CIEMAT, Laboratorio de Metrologa de Radiaciones Ionizantes,
Avda. Complutense, 22, 28040, Madrid, Espana.
Recibido el 10 de marzo de 2010; aceptado el 31 de agosto de 2010
With the Bonner spheres spectrometer neutron spectrum is obtained through an unfolding procedure. Monte Carlo methods, Regularization,
Parametrization, Least-squares, and Maximum Entropy are some of the techniques utilized for unfolding. In the last decade methods based
on Artificial Intelligence Technology have been used. Approaches based on Genetic Algorithms and Artificial Neural Networks have been
developed in order to overcome the drawbacks of previous techniques. Nevertheless the advantages of Artificial Neural Networks still it has
some drawbacks mainly in the design process of the network, vg the optimum selection of the architectural and learning ANN parameters.
In recent years the use of hybrid technologies, combining Artificial Neural Networks and Genetic Algorithms, has been utilized to. In this
work, several ANN topologies were trained and tested using Artificial Neural Networks and Genetically Evolved Artificial Neural Networks
in the aim to unfold neutron spectra using the count rates of a Bonner sphere spectrometer. Here, a comparative study of both procedures has
been carried out.
Keywords: Neutron spectrometry; neural networks; evolutive algorithms.
Con el espectrometro de esferas Bonner se puede obtener el espectro a traves de un procedimiento de reconstruccio n. Los metodos Montecarlo, de Regularizacien, de parametrizacion, de mnimos cuadrados, de la maxima entropa son algunas de las tecnicas utilizadas para la
reconstruccion. En la u ltima decada, se han utilizado los metodos basados en la tecnologa de Inteligencia Artificial. Se han desarrollado
metodos basados en Algoritmos Geneticos y Redes Neuronales Artificiales en un intento de resolver las desventajas de las tecnicas mencionadas. Sin embargo, a pesar de la ventajas de las redes neuronales, las mismas presentan algunos inconvenientes principalmente en lo
que se refiere al proceso de dieno de de las redes, por ejemplo, la seleccion o ptima de los parametros de arquitectura y aprendizaje. En
anos recientes, tambien se ha utilizado tecnologas hbridas, combinando las redes neuronales y los algoritmos geneticos. En e ste trabajo, se
disenaron y entrenaron varias topologeas de redes neuronales y redes neuronales evolucionadas geneticamente con el objetivo de reconstruir
espectros de neutrones utilizando las tasas de conteo de un espectrometro de esferas Bonner. Aqu se realiza un estudio comparativo de
ambos procedimientos.
Descriptores: Espectrometra de neutrones; redes neuronales; algoritmos evolutivos.
PACS: 29.30.Hs; 29.40-n; 07.05-Mh

Introduction

Neutron energy spectra found in workplaces are often complex, the range of neutron energies involved can extend over
nine or ten orders of magnitude. To improve the assessment
of personal equivalent dose (Hp) and ambient dose equivalent (H * (10)) in workplace, requires the proper characterization of neutron spectra [1]. The monitoring of occupational radiation exposure in neutron fields is mainly done
with multi-element systems where each element has a particular response to neutrons. With the neutron spectrum information and neutron fluence-to-dose conversion coefficients,
different dose quantities, like Hp or H * (10), can be estimated.

With the count rates taken with the Bonner Spheres Spectrometer (BSS), the neutron spectrum can be unfolded. In
BSS, each detector is characterized by a response function,
the whole set is the response matrix. The relationship among
neutron spectrum, count rates and response matrix is described by the integral-differential equation of Fredholm of
first type. Due to the number of detectors is smaller than
the number of energy groups used to describe the spectrum,
resulting problem is ill-conditioned therefore, unfolding procedures should be applied [1]. In previous works, have been
reported neutron spectrometry and dosimetry results, by using the Artificial Intelligence (AI) technology as alternative
solution. The Artificial Neural networks (ANN) [2] have re-

J.M. ORTIZ-RODRIGUEZ et al.

ceived the higher attention among researchers [1], however,

the use of this technology is not free of problems [3].
ANN technology is a useful alternative to solve the neutron spectrometry problem, however, several drawbacks must
be solved in order to simplify the use of these procedures.
Many of the previous studies in neutron spectrometry and
dosimetry by using the ANN approach, have found serious drawbacks in the ANN design process itself, mainly in
the proper determination of the structural and learning parameters of the networks being designed. These parameters
are significant contributing factors to the ANN performance;
however, the optimal selection follows in practical use no
rules because they are generally heuristically chosen by using
the trial and error technique, which produces networks with
poor performance and low generalization capacity. For the
anterior, the nuclear research community needs approaches
that implement ANN models faster than what is currently
available. In consequence, more research has been suggested
in order to overcome these drawbacks.
Recently, the use of ANN technology has been applied
with success in the neutron spectrometry and dosimetry problems, using a novel approach known as Robust Design of Artificial Neural Networks (RDANN) methodology in the design process of the networks [3]. Another promising technique for ANN design, is to introduce the capacity of adaptation to the net by using Evolutionary Algorithms (EA) [4],
which can be used to adapt the connections among the synaptic weights, the design of the architecture, the adaptation
of the learning rule, etc. The use of a hybrid technology
based in the RDANN methodology combined with EA for
the modelling of ANN applied in the neutron spectrometry
and dosimetry problems could be a very convenient alternative technique. In this work, several ANN topologies were
trained and tested using two ANN design approaches, to and
the optimum parameters of ANN. The networks were designed in the aim to unfold neutron spectra using the count
rates of CIEMAT BSS, Spain. Here, a comparative study of
both procedures has been carried out.

2. Materials and methods

ANN and EA are two relatively young research areas that
were subject to a steadily growing interest during the past
years. Until today, many researchers still prefer use the
gradient search method Back Propagation (BP) in training
ANN [2]. However, this technique is a local search method
and when applied to complex nonlinear optimization problems, can sometimes result in inconsistent and unpredictable
performances. One of the main hindrances is due to the fact
that searching of optimal weights is strongly dependent on
initial weights and if they are located near local minima, the
algorithm would be trapped; if the initial guess of the ANN
is near local maxima, it will climb the gradient and get stuck.
The structure of a neural network is a significant contributing
factor to its performance and the structure is generally heuristically chosen. Several different attempts have been proposed

by various researchers to alleviate the training problems. The

use of EA as search technique has allowed different properties of ANN to be evolved.
In this work, the performance of neural nets, designed
for the BSS of CIEMAT, with two desing methodologies
is compared. The traditional nets (ANN) were designed
by means of the RDANN methodology and the Evolutive
ANN (EANN) [4], were designed by using the computer
program known as NeuroGenetic Optimizer (NGO) [5]. To
use the knowledge stored at the networks designed with both
metodologies, friendly computer programs whit graphical interfaces were designed and builded respectively. A modified
version of the NSDann unfolding code was used for the traditional ANN [6], and the unfolding code called Neutron
Spectrometry and Dosimetry based on Evolutive Artificial
Neural networks (NSDEann) was designed for EANN.
In both cases, the first layer (input) corresponds to the
12 Bonner spheres of the CIEMATs BSS, the hidden layers and neurons by layer should be determined and the third
layer (output) has 85 neurons. The first 72 outputs correspond
to the neutron spectra and the remaining 13 are equivalent
doses. In both methodologies training was carried using 201
spectra, and 50 spectra were used for testing the learning of
the nets.
Utilizing the RDANN methodology, additional parameters have to be determined: momentum, learning rate, training algorithm and mean square error (MSE). The EANN
technology, although does not requires on the part of the user
ANN parameters to be determined, it requires the determination of parameters regarded with the GA used to and the
architecture of the net.
This is a very difficult problem, which is similar to that
found in ANN design and, at present, this represents a serious drawback of the EANN approach, because the parameter
selection is done through the trial and error technique. More
research is needed in this sense. At present, work is being
carried out. The neutron spectra unfolded and its corresponding 13 equivalent doses were computed with the modified
NSDann and NSDEann unfolding codes, for a 239 PuBe neutron source measured at 80 cm distance, at Departamento
de Ingeniera Nuclear (DIN) of Universidad Politecnica de
Madrid (UPM).

Results and discussion

By means of the RDANN methodology, 36 traditional ANN

topologies were designed and trained, each in a time of
107 seconds average. The optimum network topology is: 12
neurons in the input layer, 16 nerons in a hidden layer and 85
in the output layer. The neural nets were trained until MSE
was reduced to 104 . Additional parameters are: momentum: 0.001, learning rate: 0.1, training algorithm: trainscg.
After the training and testing stages, the knowledge stored at
synaptic weights was extracted and a modified version of the
NSDann unfolding code was designed. Figure 1, shows the

Rev. Mex. Fs. S 57 (1) (2011) 8992

PERFORMANCE OF ARTIFICIAL NEURAL NETWORKS AND GENETICAL EVOLVED ARTIFICIAL NEURAL NETWORKS. . .

F IGURE 1. Neutron spectra unfolded and equivalent doses calculated with NSDann.

F IGURE 2. Neutron spectra unfolded with NSDEann.

neutron spectra unfolded and its 13 equivalent doses calculated by means of the modified NSDann code.
By using the EANN methodology, 600 EANN were designed and trained in a time of 04:44:09. The minimum network training passes for each network were 420, the cutoff
for network training passes was 450, and the limit on hidden neurons was 8. The optimum network topology is: a
Fast-Back Propagation neural network with 12 inputs, 6 logistic neurons in a hidden layer and 85 neurons in the output
layer, with an accuracy on training set= 99.50% according
NGO. The parameters used for the GA, selected using the
trial and error technique, are: generations run: 10, population
Size: 60, selection was performed by the top 50% surviving,
re?lling of the population was done by cloning the survivors,
mating was performed by using the TailSwap technique, mutations were performed using Random Exchange technique
at a rate of 25%.
After the training and testing stages, it was not possible
to extract the knowledge stored at synaptic weights of the
network designed with NGO. For this reason, the NSDEann
unfolding code was designed to read and graph the spectrometric and dosimetric information produced by means of the
EANN trained with NGO. Figure 2, shows the neutron spectra unfolded and its 13 equivalent doses calculated by means
of the modified NSDann.
In order to compare the results obtained with the NSDann
and NSDEann unfolding codes, the computer tool known as
Neutron spectrometry and dosimetry Tool Box (NSDTB)
was used [7]. Figure 3, shows the performance of the neutron spectra unfolded with the EANN approach (green line)
and the spectra unfolded with the RDANN methodology (red
line) and its corresponding 13 equivalent doses. As can be
seen, the spectra and doses present diferencies, this is due
mainly to the trial and error approach used to select the GA
parameter in the EANN approach, which is a serious drawback of this method. More research is needed in this sense.
At present work is being developed to overcome these drawbacks.
From Fig. 3, can be observed that the neutron spectra unfolded and equivalent doses calculated with both, traditional
ANN and EANN, are very similar, however, the EANN unfold the spectra in different energy bins, this makes that the
2 test fails. A deeper analysys should be needed in order to
observe closer what could be happening.

F IGURE 3. Neutron spectra unfolded with NSDann - NSDEann.

Conclusions

The use of ANN technology is a useful alternative to solve

Rev. Mex. Fs. S 57 (1) (2011) 8992

J.M. ORTIZ-RODRIGUEZ et al.

CIEMAT BSS system. The neural net was trained and tested
using a large set of neutron spectra compiled by the IAEA.
The success of ANN technology in neutron spectrometry and dosimetry, using only the Bonner spectrometer count
rates as input in the trained network will overcome some
of the problems associated with the solution of such illconditioned problem. The results here reported demonstrate
that the use of this technology has become in a useful tool.
Until now RDANN methodology seems to be mo reliabe
in the neutron spectra unfoloding problem, because more research has been carried out in this sense and consecuently
more information is available. The main drawback of EANN
is the trial and error technique used to determine the optimum
values of the GA used to build the network. The combination

of RDANN whit GA could improve widely the results obtained with both methodologies.
More research is needed and at present, work is being
carried out.

1. H.R. Vega-Carrillo, V.M. Hernandez-Davila, E. ManzanaresAcuna, E. Gallego, A. Lorente, and M.P. Iniguez, Radiation
Protection Dosimetry 126 (2007) 408.

4. J.M. Ortiz-Rodrguez, M.R. Martnez-Blanco, E. Gallego Daz

and H.R. Vega-Carrillo, Electronics, Robotics and Automotive
Mechanics Conference, 4 (2008) 387.

2. S. Haykin, Neural Networks: A comprehensive foundation.

Prentice Hall Inc. (1999).

6. M.R. Martnez-Blanco, J.M. Ortiz-Rodrguez, and H.R. VegaCarrillo, Electronics, Robotics and Automotive Mechanics Conference 0 (2009) 131.

3. J.M. Ortiz-Rodrguez, M.R. Martnez-Blanco, and H.R. VegaCarrillo, Electronics, Robotics and Automotive Mechanics Conference 2 (2006) 131.

7. J.M. Ortiz-Rodrguez, M.R. Martnez-Blanco, E. Gallego Daz,

and H.R. Vega-Carrillo, Electronics, Robotics and Automotive
Mechanics Conference 5 (2009) 112.

Acknowledgements
This work was supported by Agencia Espanola de Cooperacion Internacional para el Desarrollo, AECID (Spain).

5. NeuroGenetic Optimizer. http://www.biocompsystems.com

Rev. Mex. Fs. S 57 (1) (2011) 8992

J Radioanal Nucl Chem (2009) 281:615618

DOI 10.1007/s10967-009-0057-z

Spectra and dose with ANN of

252

Cf,

241

AmBe, and

239

PuBe

Hector Rene Vega-Carrillo M. R. Martinez-Blanco

Victor Martin Hernandez-Davila
Jose Manuel Ortiz-Rodriguez

Received: 16 April 2009 / Published online: 4 September 2009

Akademiai Kiado, Budapest, Hungary 2009

Abstract Artificial neural networks have been applied to

unfold the neutron spectra and to calculate the effective
dose, the ambient equivalent dose, and the personal dose
equivalent for 252Cf, 241AmBe, and 239PuBe neutron
sources. The count rates that these neutron sources produce
in a Bonner Sphere Spectrometer with a 6LiI(Eu) were
utilized as input in both artificial neural networks. Spectra
and the ambient dose equivalent were also obtained with
BUNKIUT code and the UTA4 response matrix. With both
procedures spectra and ambient dose equivalent agrees in
less than 10%. The Artificial neural network technology is
an alternative procedure to unfold neutron spectra and to
perform neutron dosimetry.
Keywords Spectrum Dose Isotopic neutron source
Artificial neural networks

several hundreds of MeV. Also, they are in a broad variety

of energy distributions, named neutron-fluence spectrum or
simply neutron spectrum, UE(E) [2, 3].
Isotopic neutron sources are utilized in several applications that requires to know the neutron spectrum and the
dose [46]. Some of the isotopic neutron sources are the
252
Cf, 241AmBe, 241AmB, 239PuBe, 238PuBe, etc.
The Bonner sphere spectrometer, BSS, or multisphere
neutron spectrometer, is a set of high-density polyethylene
spheres with a thermal neutron detector in their centre that
is utilized to obtain, through an unfolding process, the
UE(E). Each sphere-detector combination has a unique
response, the whole set of responses is the response matrix,
RU(E).
Unfolding process involves solving Eq. 1, which is a
Fredholm integral equation of the first kind.
Z
Ci RU;i EUE EdE
1

Introduction
The measurement of the intensity of a radiation field with
respect to certain quantity like angle, energy, frequency,
etc., is very important in radiation spectrometry having, as
a final result, the radiation spectrum [1, 2].
Neutrons are found in the environment or are artificially
produced by different ways; these neutrons have a wide
energy range extending from few thousandths of eV to

H. R. Vega-Carrillo (&) M. R. Martinez-Blanco

V. M. Hernandez-Davila J. M. Ortiz-Rodriguez
Unidad Academica de Estudios Nucleares de la UAZ,
Apdo. Postal 336, 98000 Zacatecas, Zac., Mexico
e-mail: fermineutron@yahoo.com

Here, Ci is the count rate measured with the ith spheredetector combination. Response functions have poor energy
resolution and are not fully linearly independent of each
other. The spectrum is defined in a large number of energy
groups in comparison to the number of spheres used in the
spectrometer thus; Equation 1 is an ill-conditioned problem
[7]. Total fluence is obtained integrating the UE(E) for all
energies as shown in Eq. 2.
Z
/ UE EdE
2
Once the UE(E) is obtained the required dose, D, is
obtained using Eq. 3.
Z
D dU EUE EdE
3

123

616

Here, dU(E) are the fluence-to-required dose conversion

coefficients [7]. Therefore, the main problem is to solve
Eq. 1 and several procedures are utilized [8].
In the brain, a neural network is a massively parallel
distributed processor with a natural propensity for saving
experiential knowledge, previously acquired through a
learning process, making it available for latter use. The
Artificial Neural Networks, ANNs, have been proposed to
emulate this feature from brains behaviour [9].
The mathematical models used to design an ANN have
at least three layers, the input, hidden and the output. Each
layer has several process units, the neurons, which are
connected through synaptic weights where the knowledge
is stored. The activation of a neuron depends on the given
weight of the synapses that are appraised through the
activation functions [10, 11].
The ANN learning is the adaptation process to the
training data. The training has the purpose of selecting the
weights that adapts better to the network in relation to
training data. During ANN training a set of data, with input
and output information, is utilized and the synaptic weights
are adjusted until, at some point, the pattern between input
and output data is accomplished.
The ANN technology has been utilized to unfold the
neutron spectrum, and with this the neutron dose has been
estimated [7, 10, 11]. Also, The ANNs have been applied
to obtain directly the neutron doses without the need of
neutron spectrum information [8]. In all these applications
only the count rates obtained with a BSS are required.
In this work the ANN technology has been utilized to
unfold the neutron spectrum and to determine the neutron
doses of 252Cf, 239PuBe, and 241AmBe neutron sources,
using the count rates measured with a BSS. These results
were compared with spectra and the H*(10) obtained with
BUNKIUT code.

Materials and methods

Two different ANNs were designed and trained, the first to
perform the neutron spectrometry and the second to calculate 13 dosimetric quantities. The lethargy spectra, taken
from the compilation made by the International Atomic
Energy Agency [12], were converted to energy spectra.
With the MCNP 4C code [13], the spectra were re-binned
from their original energy distribution defined in 60 energy
groups to 31 energy groups [14], which is the energy
structure utilized in BUNKIUT code [15]. For each spectrum
the EAP, EPA, ERLAT, ELLAT, EROT, EISO, H*(10), Hp,s(10,0),
Hp,s(10,15), Hp,s(10,30), Hp,s(10,45), Hp,s(10,60),
Hp,s(10,75) were calculated using the fluence-to-dose
conversion coefficients from ICRP 74 [16].

123

H. R. Vega-Carrillo et al.

The re-binned spectra were utilized, together with

UTA4 matrix response, [17] to calculate their respective
count rates produced in the Bonner spheres whose diameters are 0, 5.08, 7.62, 12.7, 20.32, 25.4 and 30.48 cm.
One hundred seventy-seven neutron spectra, their doses
and their respective BSS count rates were utilized to train
and test both ANNs. From the available set of data the 10%
was randomly selected and isolated from the whole set to
test the ANN, the remainder was left for training purposes.
Both ANNs were trained and tested using MATLAB
software.1 Topology of ANN for spectrometry was 7:140:
500:140:70:27, where each number is the amount of neurons in each layer. For dosimetry the ANN topology was
7:140:200:401:140:13. Both ANNs were feed-forward with
backpropagation algorithm and variable learning rate [18].
After training the ANNs performance was verified using
the data set that was not utilized for training purposes.
The BSS was utilized to obtain the count rates produced
by 252Cf, 239PuBe and 241AmBe. Their respective activities were 4.34 9 108, 1.85 9 1011 and 3.7 9 109 Bq. During BSS measurements the source-to-spectrometer distance
was 120, 100 and 50 cm for 252Cf, 239PuBe and 241AmBe,
respectively. The count rates were utilized to unfold the
neutron spectra using BUNKIUT code and UTA4 response
matrix; the neutron spectra was utilized to estimate the
H*(10). Neutron spectra were also unfolded with the ANN
designed for spectrometry, and the dosimetric quantities
were calculated using the ANN designed for dosimetry.

Results and discussion

In Fig. 1 is shown the neutron spectrum of 252Cf obtained
with BUNKIUT code and the ANN for spectrometry. Here
can be noticed that both results have good agreement. The
total fluence is 193 3 and 191 3% cm-2 s-1, for
spectrum obtained with BUNKUT and the ANN, respectively. The uncertainty was taken from the uncertainty of
count rates, therefore in both cases no uncertainties were
assigned to the unfolding process.
In Fig. 2 is shown the 239PuBe neutron spectrum
unfolded with BUNKIUT and the spectrum unfolded with
the ANN trained to perform neutron spectrometry. The
shape of both spectra looks alike. The total fluence is
121 5 and 128 5% cm-2 s-1, for spectrum obtained
with BUNKUT and the ANN, respectively. Uncertainties
were assigned from those from BSS count rates and no
uncertainties were assigned to the unfolding process.
The 241AmBe neutron spectra obtained with BUNKIUT and the ANN for spectrometry is shown in Fig. 3. It
1

MATLAB is a trade mark of The MathWorks, Inc.

252

Spectra and dose with ANN of

Cf,

241

AmBe, and

239

PuBe

617
400

350

BUNKIUT
ANN

300

250
200
150

239
241

10 2

Neutron energy [MeV]

Fig. 1 Neutron spectrum of

252

Cf source

Fig. 4 Dosimetric quantities for

Lethargy fluence [ cm-2 - u-1]

BUNKIUT
ANN

10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

101

102

10 0

10 1

10 2

Neutron energy [MeV]

Fig. 2 Neutron spectrum of

239

PuBe source

3,0

Lethargy fluence [cm -2-u -1 ]

241

2,5

AmBe
BUNKIUT
ANN

2,0

1,5

1,0

0,5

0,0
10 -8

10 -7

10 -6

10 -5

10 -4

10 -3

10 -2

10 -1

Neutron energy [MeV]

Fig. 3 Neutron spectrum of

239

PuBe,

241

AmBe and

252

PuBe

10-8

PuBe
AmBe
Cf

Dosimetric quantity

50
239

Hp(10,30o)

10 1

Hp(10,15o)

10 0

H*(10)

10 -1

Hp(10,0o)

10 -2

Erot

10 -3

Eiso

10 -4

Ellat

10 -5

Erlat

10 -6

Epa

10 -7

Eap

100

0
10 -8

252

Hp(10,75o)

Hp(10,60o)

Hp(10,45o)

252

Dose [pSv-cm2]

Lethargy neutron fluence [cm-2 - u-1]

241

AmBe source

can be noticed that both are similar. The total fluence for
both spectra is 8 5% cm-2 s-1. As in both previous
cases no uncertainties were assigned to the unfolding
process and the 5% comes from the BSS count rates.

The presence of epithermal and thermal neutrons in the

spectra is due the room return effect [19]. In all the cases
the ANN for spectrometry was able to unfold the neutron
spectra using only the BSS count rates.
The dosimetric quantities obtained with the ANN for
dosimetry are shown in Fig. 4. The neutron fluenceto-ambient dose equivalent factors, h*(10), obtained with the
ANN were 282, 270, and 322 pSv-cm2. With BUNKIUT
these values were 296, 265 and 314 pSv-cm2 for 239PuBe,
241
AmBe, and 252Cf, respectively. The differences go
from 1.7 to 5%. According to IAEA [12] the h*(10) for
241
AmBe is 395 pSv-cm2 while for 252Cf is 380 pSv-cm2,
these values are assuming point-like sources in vacuum. In
this work h*(10) was larger for 252Cf than 241AmBe due
to the changes in the neutron spectra due to experimental
conditions.
The H*(10) calculated with the ANN were 3.593 9 104,
2.272 9 103, and 6.153 9 104 pSv, with BUNKIUT these
values are 3.573 9 104, 2.192 9 103, and 6.050 9 104
pSv for 239PuBe, 241AmBe and 252Cf, respectively. The
differences between those values go from 0.6 to 3.6%.
Considering the uncertainties of the BSS count rates, the
differences noticed in h*(10) and H*(10) are not relevant.
The ANN for dosimetry was able to calculate the neutron
doses using only the measured BSS count rates.

Conclusions
Two ANNs, one for spectrometry and other for neutron
dosimetry, were utilized to unfold the neutron spectra and to
estimate the dosimetric features of, 252Cf, 239PuBe, and
241
AmBe neutron sources. Spectra and doses were also
calculated using the BUNKIUT code. Results obtained with
both procedures are in agreement. The ANN technology has

123

618

been successfully applied utilizing as input the BSS count

rates. Therefore, this technology is an alternative procedure
for neutron spectrometry and dosimetry.
Acknowledgments This work is part of project SYNAPSIS partially supported by CONACyT (Mexico) under contract SEP-2004C01-46893.

References
1. Thomas DJ, Klein H (2003) Introduction. Radiat Prot Dosim
107:13
2. ICRU (2001) Determination of operational dose equivalent
quantities for neutrons. J ICRU 1:17
3. Alevra AV (1999) Neutron spectrometry. Radioprotection 34:305
4. DiPrete DP, DiPrete CC, Sigg RA (2008) Neutron activation
analysis applications at the Savannah River site using an isotopic
neutron source. J Radioanal Nucl Chem 277:43
5. Griffin MA, Landsberger S, Biegalski S, Cournoyer ME (2008)
Characterization of a Plutonium Beryllium source for use in
neutron damage studies. J Radioanal Nucl Chem 276:807
6. Turhan S (2007) Efficiency calibration o fan HPGe detector in the
0.1-2.5 MeV energy range for Am-Be neutron source-based
PGAA applications. J Radioanal Nucl Chem 273:443
7. Vega-Carrillo HR, Hernandez-Davila VM, Manzanares-Acuna E,
Mercado GA, Iniguez MP, Barquero R, Palacios F, Mendez R,
Arteaga T, Ortiz JM (2006) Neutron spectrometry using artificial
neural networks. Radiat Meas 41:425
8. Vega-Carrillo HR, Hernandez-Davila VM, Manzanares-Acuna E,
Mercado GA, Gallego E, Lorente A, Perales-Munoz WA, RoblesRodriguez JA (2006) Artificial neural networks in neutron
dosimetry. Radiat Prot Dosim 118:251

123

H. R. Vega-Carrillo et al.
9. Haykin S (1994) Neural networks: a comprehensive foundation,
1st edn. Macmillan College Publishing Co, Englewood Cliffs, NJ
10. Cordes E, Fehrenbacher G, Schutz R, Sprunck M, Hahn K,
Hofmann R, Biersack JP, Wahl W (1998) An approach to unfold
the response of a multi-element system using an artificial neural
network. IEEE Trans Nucl Sci 45:1464
11. Kardan MR, Setayeshi S, Koohi-Fayegh R, Ghiassi-Nejad M
(2003) Neutron spectra unfolding in Bonner spheres spectrometry
using neural networks. Radiat Prot Dosim 104:27
12. IAEA (2001) Compendium of neutron spectra and detector
responses for radiation protection purposes. Supplement to
Technical Reports Series No. 403. Vienna
13. Briesmeister, JF (ed) (2000) MCNPTM: a general Monte Carlo
N-particle transport code. Los Alamos National Laboratory
Report LA-13709-M. Los Alamos, NM
14. Vega-Carrillo HR, Manzanares-Acuna E, Ortiz-Rodriguez JM,
Arteaga-Arteaga T (2007) Neutron spectra re-binning and dose
calculation using Monte Carlo methods. Revista Mexicana de
Fisica S 53:1
15. Lowry KA, Johnson TL (1984) Modification to iterative recursion
unfolding algorithm and computer codes to find more appropriate
neutron spectra. Naval Research Laboratory Memorandum 5340,
Washington, DC
16. ICRP (1996) Conversion coefficients for use in radiological
protection against external radiation. ICRP 74. Ann ICRP 26:199
17. Hertel NE, Davidson JW (1985) The response of Bonner spheres
to neutrons from thermal energies to 17.3 MeV. Nucl Instrum
Methods A 238:509
18. Vega-Carrillo HR, Hernandez-Davila VM, Manzanares-Acuna E,
Gallego E, Lorente A, Iniguez MP (2007) Artificial neural networks technology for neutron spectrometry and dosimetry. Radiat
Prot Dosim 126:408
19. Vega-Carrillo HR, Manzanares-Acuna E, Iniguez MP, Gallego E,
Lorente A (2007) Study of isotopic neutron sources inside concrete walls spherical cavities. Radiat Meas 42:1373

REVISTA MEXICANA DE FISICA S 56 (1) 3539

FEBRERO 2010

Different spectra with the same neutron source

H.R. Vega-Carrilloa, , J.M. Ortz-Rodrgueza,b , V.M. Hernandez-Davilaa,b ,
Ma. del R. Martnez-Blancoa , B. Hernandez-Almaraza , and A.A. Ortz-Hernandeza
a
Unidad Academica de Estudios Nucleares Universidad Autonoma de Zacatecas,
Apartado Postal 336, Zacatecas, 98000, Zac. Mexico,
Telephone: (492) 922 7043, Ext.118 and 119,

e-mail: fermineutron@yahoo.com
b
Departamento de Electrotecnia y Electronica, Escuela Politecnica Superior,
Avda. Menendez Pidal s/n Cordoba, Spain.
G.A. Mercado
Unidad Academica de Matematicas, Universidad Autonoma de Zacatecas,
Jardn Juarez 147, 98000 Zacatecas, Zac. Mexico.
Recibido el 11 de marzo de 2009; aceptado el 11 de agosto de 2009
Using as source term the spectrum of a 239 Pu-Be source several neutron spectra have been calculated using Monte Carlo methods. The
source term was located in the centre of spherical moderators made of light water, heavy water and polyethylene of different diameters. Also
a 239 Pu-Be source was used to measure its neutron spectrum, bare and moderated by water. The neutron spectra were measured at 100 cm
with a Bonner spheres spectrometer. Monte Carlo calculations were used to calculate the neutron spectra of bare and water-moderated spectra
that were compared with those measured with the spectrometer. Resulting spectra are similar to those found in power plants with PWR, BWR
and CANDU nuclear reactors. Beside the spectra the dosimetric features were determined. Using moderators and a single neutron source
can be produced neutron spectra alike those found in workplaces, this neutron fields can be utilized to calibrate neutron dosimeters and area
monitors.
Keywords: Neutron sources; neutron transport; Monte Carlo applications.
Mediante el espectro de una fuente de 239 Pu-Be varios espectros de neutrones han sido calculados mediante metodos Monte Carlo donde el
termino fuente se coloco en el centro de moderadores esfericos de agua ligera, agua pesada y polietileno de diferentes diametros. Una fuente
real de 239 Pu-Be se utilizo para medir el espectro de neutrones mediante un espectrometro de Esferas Bonner. El espectro se midio a 100 cm
de distancia de la fuente desnuda y de la fuente dentro de un moderador cilndrico de agua. Con la informacion de los espectros se obtuvieron
las caractersticas dosimeticas de la fuente desnuda y de la moderada con agua. Mediante calculos Monte Carlo se calcularon los espectros y
se compararon con los espectros medidos. Los espectros resultantes son similares a los reportados en plantas nucleares con reactores BWR,
PWR y tipo CANDU. Mediante el uso de moderadores se pueden producir espectros que se asemejan a situaciones reales y se pueden usar
para calibrar dosmetros y monitores de a rea.
Descriptores: Fuente de neutrones; transporte de neutrones; aplicaciones Monte Carlo.
PACS: 29.25.Dz; 28.20.Gd; 87.53.Wz

Introduction

Neutron energy spectra found in workplaces are often complex, the range of neutron energies involved can extend over
nine or ten orders of magnitude. To improve the assessment
of personal equivalent dose (Hp ) and ambient dose equivalent [H*(10)] in workplace requires the proper characterization of neutron spectra. Dosimeters used to survey either the
personal equivalent or the ambient equivalent doses have responses that strongly vary with neutron energy [1,2].
During calibration of dosimetric devices utilized in neutron fields the use of calibrating sources with spectra alike to
those meet in practice is strongly recommended [3-5].
To calibrate dosimetric devices the International Organization for Standardization (ISO) recommends 252 Cf
and 252 Cf/D2 O, 241 AmBe, and 241 AmB as calibration
sources [6]. However, these source have spectra that are different from those found in actual situations [7,8].

To produce realistic neutron spectra whose spectrum look

alike to those found in workplaces different approaches have
been made in several facilities [1,2,8-10].
The neutron spectrum of a primary neutron source is
modified using different materials, some of which are moderators, this practice is commonly used to obtain a desired
neutron spectra, where beside to produce a realistic neutron
spectra, the dosimetric features must be determined [11].
In aim to this work was to produce different neutron spectra using the same neutron source. To achieve this objective
the neutron spectra of a 239 Pu-Be neutron source was modified by inserting the neutron source in the centre of moderating materials, this part was done using Monte Carlo calculations. Also, the neutron spectrum of a 239 Pu-Be source, bare
and water-moderated, was measured and calculated; with the
spectra information the dosimetric features were calculated.
In this case the source and the moderator are cylindrical.

H.R. VEGA-CARRILLO ET AL.

the Bonner spheres spectrometer. Height of water moderator

was 30 cm. The bare and water-moderated spectra of 239 PuBe source were unfolded using the BUNKIUT code and the
UTA4 response matrix [13]. Resulting spectra, E (E), were
utilized to estimate the total fluence, , and the ambient dose
equivalent, H*(10), using the ICRP 74 fluence-to-dose conversion coefficients, h*(10) [14]; these magnitudes were calculated using Eqs. (1) and (2).
Z
= E (E)dE
(1)
E

H (10) =

E (E) h (10)dE

(2)

F IGURE 1. Ball 2-moderated 239 Pu-Be spectra.

Nowadays the area neutron monitor Berthold LB 6411

is widely utilized to survey the H*(10); the response of this
device is not exactly the same of ICRP 74. Thus the bare
and the water-moderated 239 Pu-Be spectra were utilized to
calculate the H*(10) using the response function of Berthold
LB 6411 [15].

3.
3.1.

F IGURE 2. Ball 5-moderated 239 Pu-Be spectra.

2. Materials and methods

2.1. Spherical moderators
The neutron spectra produced by point-like 239 Pu-Be isotopic
neutron source located inside spherical moderators were calculated using the MCNP code [12]. Moderating spheres were
5.08, 7.62, 12.7, 25.4, and 30.8 cm in diameter and were
modelled as water, heavy water and polyethylene were used
as moderating materials. Neutron spectra were calculated at
100 cm from the centre of the spheres, were the point-like
source term was located. The amount of histories utilized
for each moderator was large enough to have a Monte Carlo
uncertainty of 5%.
2.2. Cylindrical moderator
The neutron spectrum produced by a 239 Pu-Be at 100 cm in
an open space at 200 cm above floor level was measured using the Bonner spheres spectrometer with 0.4 0.4 mm
6
LiI(Eu) scintillator. The source is 1.851011 Bq, is distributed in three cylindrical pellets 2.54 14 cm. Then
the source was inserted in a 28 40 cylindrical container
with water, and the neutron spectrum was measured with

Results and discussion

Spherical moderators

A neutron source spectrum can be modified with the use of

moderating materials; modified spectrum depend upon the
size and type of moderator. In Fig. 1 are shown the spectra
of a 239 Pu-Be point-like neutron source located in the centre of a 5.08 cm-diameter sphere (Ball 2) made of water,
heavy water and polyethylene. Resulting spectra from water and polyethylene moderators are alike, having epithermal
and thermal neutrons. It can be noticed that moderating features of polyethylene are better than those for water, probably
due to differences in their absorption features. The spectrum
produced by the heavy water moderator is quite similar to the
bare 239 Pu-Be source because the amount of heavy water in
the Ball 2 container is small to produce strong effect in the
spectrum.
In Fig. 2 the resulting spectra of a point-like 239 PuBe located in the centre of 12.7 cm-diameter moderator (Ball 5)
of water, heavy water and polyethylene are shown. As in the
case of Ball 2 moderators, water and polyethylene moderators have similar spectra; the peak in the 1 to 10 MeV tend to
decrease while the thermal peak tend to increase as the moderator size is increased. The spectrum produced in the heavy
water moderator has a large contribution of epithermal neutrons in the region from 103 to 0.4 MeV in comparison with
the amount of neutrons with energy less than 1 keV.
In Fig. 3 moderated 239 Pu-Be spectra are shown for
25.4 cm-diameter moderators (Ball 10). Here, the spectrum
produced by the polyethylene moderator is smaller than the
spectrum produced by the water moderator, but the shapes are
very alike. On the other hand the spectrum produced by the
heavy water moderator is quite similar to the spectrum produced by the heavy water moderated 252 Cf source [6], even

Rev. Mex. Fs. S 56 (1) (2010) 3539

DIFFERENT SPECTRA WITH THE SAME NEUTRON SOURCE

when the unperturbed spectrum of 239 Pu-Be is quite different of the unperturbed 252 Cf. From this figure is evident that
heavy water moderator absorbs less neutrons that other two
moderators.
In Fig. 4 is shown the neutron spectra of a point-like
Pu-Be source located inside a 30.48 cm-diameter moderators (Ball 12). The tendency observed for the ball 10 is kept
for the ball 12. Spectrum produced in the water moderator
is alike to spectrum produced by the polyethylene moderator,
where this last produces less neutrons.
239

Using Eq. (1) the total neutron fluence produced by the

different spherical moderators were calculated. In Fig. 5
is shown the total neutron fluence produced by the moderators. The total fluence produced by the heavy water moderators is almost constant regardless the moderator diameter, meaning that neutron capture in heavy water is negligible in comparison to water and polyethylene moderators.
For water and polyethylene spherical moderators, with diameters larger than 12.7 cm, the total fluence produced by
polyethylene moderators is smaller than total fluence produced in water moderators, this is due to differences in the
neutron capture and scattering features of both moderators.
This is an important fact that must to be taken into account for
safety reasons, because neutron capture can produce prompt
gamma-rays that could affect the response of instrumentation
and safety of personal.
In Figs. 1 to 4 all spectra show a peak between 1 and
10 MeV. As the moderator diameter increases epithermal and
thermal neutrons increases and fast neutrons tend to decrease.
Resulting spectra for Ball 5, 10 and 12, either polyethylene
or water moderators are alike to those produced at some locations in PWR and BWR nuclear reactors [16], therefore a
239
Pu-Be inside spherical water moderators can be used to
calibrate dosimetric instruments used in this facilities.

F IGURE 4. Ball 12-moderated 239 Pu-Be spectra.

F IGURE 5. Total fluence a 100 cm produced by the moderated

Pu-Be spectra in function of moderator diameter.

239

F IGURE 6. Bonner spheres count rates produced by bare and watermoderated 239 Pu-Be source.

F IGURE 3. Ball 10-moderated 239 Pu-Be spectra.

Heavy water contains deuterium whose capture cross section is lower than hydrogen, therefore less neutrons are lost
by capture. The spectra produced by a 239 Pu-Be neutron
source inside the spherical heavy water moderators, Ball
10 and 12, is similar to neutron spectra produced by D2 ORev. Mex. Fs. S 56 (1) (2010) 3539

H.R. VEGA-CARRILLO ET AL.

moderated 252 Cf source, therefore this source can be used

in all those facilities where the 252 Cf/D2 O source is recommended for calibration. Some of the spectra produced by
the 239 Pu-Be inside the heavy water moderator are alike to
spectra reported in literature [17] for CANDU-like workplace
neutron fields.

TABLE I. The dosimetric features of bare and water-moderated

Pu-Be source at 100 cm.

239

Magnitude

Bare
239

[ cm

97 8

hEi [MeV]

3.2. Cylindrical moderator

The count rates measured with Bonner spheres at 100 cm
form the bare and water-moderated 239 PuBe source is shown
in Fig. 6, here can be noticed that for 2 and 3 inches-diameter
spheres the count rates produced by the water-moderated
source are larger than count rates produced by the bare 239 PuBe source, this is because bare source do not produce epithermal or thermal neutrons, and the response of those spheres
are large for low energy neutrons; this is also the reason why
no count rate is shown for bare detector, 0 inches-diameter
sphere. For 5 to 12 inches-diameter sphere the count rates
are larger for bare source in comparison to water-moderated
source, this is because the neutron capture, mainly in hydrogen, is increased as the neutron energy is decreased.

H*(10)ICRP 74 [ Sv h1 ]
H*(10)LB6411 [ Sv h

F IGURE 8. Water-moderated 239 Pu-Be source neutron spectrum.

Water-moderated
239

Pu-Be

64 5

5.0 0.4

2.6 0.2

141 11

59 5

132 11

55 4

With the count rates the neutron spectra were unfolded,

in Fig. 7 the bare 239 Pu-Be neutron spectra measured with
the Bonner sphere spectrometer is shown, here is also included the neutron spectra that was calculated with Monte
Carlo methods. It can be notice that both spectra are similar.
The calculated and unfolded spectra produced by 239 PuBe source moderated by water is shown in Fig. 8. The differences between calculated and measured spectra are due to
the limitations of the model of the source utilized during calculations. Uncertainties due to Monte Carlo calculations are
as large as 5% while the experimental uncertainties that are
around 10%; doing the quadratic sum of these uncertainties
gives out a overall uncertainty of 11.2%.
The dosimetric features of bare and water-moderated
239
Pu-Be source at 100 cm are shown in Table I. The neutron
mean energy of bare 239 Pu-Be is 5 MeV, this is in agreement
to value reported for this type of sources [4]. The mean energy is reduced to 2.6 MeV when the source is inserted in
the moderator of water. Moderated spectrum can be utilized
to represent the neutron spectra in power plants. If a larger
moderator is utilized the resulting spectrum could be similar
to neutron field found in medical use linear accelerators [18]
and PET cyclotrons [13,19].

4.
F IGURE 7. Bare 239 Pu-Be source neutron spectrum.

Pu-Be

Conclusions

With the use of moderating materials different neutron spectra can be produced using the same neutron source.
A precise knowledge of neutron spectral distributions of
calibrating sources is essential to have a neutron field useful
for calibration of radiation protection devices. Is no doubt
that more neutron reference sources, beside to those recommended by the ISO, are needed for the calibration of personal
neutron dosimeters and survey meters. Is highly desirable
that the spectra of those sources must be realistic to emulate
the neutron fields found in workplaces.
The feasibility of producing a neutron fields in a laboratory facility was investigated by Monte Carlo simulation.
It was found that such neutron fields could be produced by
a 239 Pu-Be neutron source located inside light water, heavy
water and polyethylene moderators.
A single isotopic neutron source in combination with water, heavy water and polyethylene moderators produce a wide
range of neutron spectra that allow to have a set of relatively

Rev. Mex. Fs. S 56 (1) (2010) 3539

DIFFERENT SPECTRA WITH THE SAME NEUTRON SOURCE

inexpensive calibrating neutron sources some of which are

alike to neutron fields found in power plants and neutron producing medical devices.
Resulting neutron spectra depend upon the size and type
of moderator; thus a bare 239 Pu-Be source looks different
when is located inside a water moderator. The mean neutron
energy varies from 5 to 2.6 MeV and the H*(10) is reduced
from 141 to 59 mSv-h1 . Using a single neutron source the
use of different moderators will lead to have neutron sources
with different dosimetric features some of which are similar
to those found in workplaces.

The spectra here calculated will be modified by the calibrating room features mainly due to room return [20], this
factor must be estimated depending of rooms size.

1. R. Bedogni, G. Gualdrini, and F. Monteventi, Nuclear Instruments and Methods in Physics Research A 476 (2002) 381.

11. N.E. Hertel and J.C. McDonald, Radiation Protection Dosimetry 35 (1991) 23-30.

2. J. Palfalvi, L. Sajo-Bohus, and M. Balasko, Nuclear Instruments and Methods in Physics Research A 476 (2002) 452.

12. J.F. Briesmeister (editor), MCNP A general Monte Carlo code

N-particle Transport Code.Los Alamos Nationai Laboratory,
Report LA-7996-M (1986).

3. ICRU, Radiation protection instrumentation and its applications. International Committee on Radiation Units Report 20
(1971).

Acknowledgements
This work is part of project SYNAPSIS that has been supported by CONACyT (Mexico) under contract SEP 2004C01-46893.

13. H.R. Vega-Carrillo, Nuclear Instruments and Methods in

Physics Research A 463 (2001) 375.

4. NCRP, Calibration of survey instruments used in radiation protection for the assessment of ionizing radiation fields and radioactive surface contamination. National Council of Radiation
Protection Report 112 (1991) 83.

14. ICRP, Conversion coefficients for use in radiological protection

against external radiation. ICRP Publication 74. Annals of the
ICRP 26 (1996) 200.

5. O.F. Naismith and B.R.L. Siebert, Radiation Protection

Dosimetry 70 (1997) 241.

15. A. Klett and B. Burgkhardt, IEEE Transactions on Nuclear Science 44 (1997) 757.

6. ISO, Neutron reference radiations for calibrating neutron measuring devices used for radiation protection purposes and for
determining their response as a function of neutron energy. International Organization for Standarization ISO 8529 (1989).
7. J.L. Chartier et al., Radiation Protection Dosimetry 61 (1995)
57.
8. R.I. Scherpelz and J.E. Tanner, Nuclear Instruments and Methods in Physics Research A 476 (2002) 400.
9. V. Gressier, G. Pelcot, J.L. Pochat, and T. Bolognese-Milstajn,
Nuclear Instruments and Methods in Physics Research A 505
(2003) 370.
10. J.C. McDonald, F. Posny, S. Gerdung-List, J.L. Chartier, and J.
Kurkdjian, Radiation Protection Dosimetry 62 (1995) 197.

16. R.V. Griffith, J. Palfalvi, and U. Madhvanath, Compendium of

neutron spectra and detector responses for radiation protection
purposes, International Atomic Energy Agency Technical Report Series No. 318 (1990).
17. J.C. Nunes, W.G. Cross, and A.J. Waker, Radiation Protection
Dosimetry 72 (1997) 11.
18. R. Barquero, R. Mendez, H.R. Vega-Carrillo, M.P. Iniguez, and
T.M. Edwards, Health Physics 81 (2005) 49.
19. R. Mendez et al., Physics Medicine and Biology 50 (2005)
5141.
20. H.R. Vega-Carrillo et al., Rev. Mex. Fs. S 54 64.

Rev. Mex. Fs. S 56 (1) (2010) 3539

Artificial Neural Networks in Spectrometry and

Neutron Dosimetry
H. R. Vega-Carrilloa, J. M. Ortiz-Rodrigueza/b, M. R. Martinez-Blancoa
and V. M. Hernandez-Davilaa/b
a

Unidad Academica de Estudios Nucleares de la Universidad Autonoma de Zacatecas, Apdo. Postal

336, 98000 Zacatecas, Zac. Mexico.
b
Escuela Politecnica Superior de la Universidad de Cordoba, Cordoba Spain

Abstract. The ANN technology has been applied to unfold the neutron spectra of three neutron
sources and to estimate their dosimetric features. To compare these results, neutron spectra were
also unfolded with the BUNKIUT code. Both unfolding procedures were carried out using the
count rates of a Bonner sphere spectrometer. The spectra unfolded with ANN result similar to
those unfolded with the BUNKIUT code. The H*(10) values obtained with ANN agrees well
with H*(10) values calculated with the BUNKIUT code.
Keywords: Artificial neural networks, neutron, spectrum, dose
PACS: 84.35.+I, 29.25.Dz, 29.30.Hs

INTRODUCTION
Neutron Spectrometry
Neutron field characterization means to define the source strength in term of a
given quantity such as energy, position, angle, etc. When energy is the variable
resulting distribution is named neutron fluence spectrum or just neutron spectrum,
E(E) [1].
Bonner sphere spectrometer, BSS, is a set of high-density polyethylene with a
thermal neutron that is located at the centre of the spheres. In Fig. 1 is shown this
spectrometer. The BSS has been widely used to determine the neutron spectrum from
thermal up to 20 MeV. To increase its response to higher energy neutrons, shells of
high-Z materials, like Pb or Cu, has been utilized in the BSS [2].
Regardless the thermal neutron detector used in BSS the shape of responses looks
alike, main difference is just noticed in the bare detector or Ball 0 due to the type of
isotope responsible of thermal neutron detection. Response amplitudes depend upon of
cross section of detecting isotope and its concentration in the detector. Each spheredetector combination has a unique response function, the full set of these responses is
the response matrix, R(E). In Fig. 2 is shown the response matrix for the BSS with
pairs of thermoluminescent dosimeters, TLDs, as thermal neutron detector [3].

Downloaded 17 Dec 2010 to 189.166.191.199. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions

FIGURE 1. Bonner sphere spectrometer.

FIGURE 2. Response matrix of a Bonner sphere spectrometer with TLDs.

Downloaded 17 Dec 2010 to 189.166.191.199. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions

Bonner sphere count rates, C, R(E) and E(E) are related through the Fredholm
integral equation of the first kind, as shown in equation 1.
C = R (E ) E (E ) dE .

(1)

Once E(E) is unfolded the desired dosimetric quantity, , is calculated; to do this

it is necessary to have the fluence-to-dose coefficients, (E), as shown in equation 2.

= (E ) E (E ) dE .

(2)

Drawbacks of BSS are the need of using an unfolding procedure to get the
spectrum, the low resolution of resulting spectra, the weight of the system, and the
excessive time required to perform the measurements; however this spectrometer is
quite reliable. Unfolding the neutron spectrum from the few count rates of BSS using
the discrete version of equation 1 mean to solve an ill-conditioned problem that has an
infinite number of solutions. Codes like BUNKIUT [4], MITOM, MAXED, FRUIT
[5], have been developed to unfold neutron spectrum from BSS count rates.

Artificial Neural Networks

In the brain, a neural network saves experiential knowledge, acquired in a learning
process, having it available for using later on. Artificial Neural Networks, ANN, has
been developed to emulate human brain [6]. An ANN has three layers of neurons: the
input, hidden and output, these are connected by synaptic weights where knowledge is
stored. The ANN learns in the training process where the synaptic weights are defined
through the activation functions. ANN training requires two set of data that are
univocally related: the input and the output data. The ANN technology have been
utilized to unfold the neutron spectrum; recent advances includes a single ANN that
gives out the neutron spectrum and the effective, E, personal, Hp, and ambient dose
equivalent, H*(10), using a user-friendly interface that do the plots and the statistical
test [7, 8].
The aim of this work is to determine the spectrum of neutrons, emitted by
spontaneous fission by a 252Cf source, as well as neutrons produced, through (, n)
reaction, by 241Am-Be and 239Pu-Be sources; also to calculate the Effective dose, the
Personal dose equivalent and the Ambient dose equivalent.

MATERIALS AND METHODS

A BSS/6LiI(Eu), with 2, 3, 5, 8, 10 and 12 inches-diameter spheres, was utilized to
measure the count rates produced by three neutron sources: 252Cf, 241Am-Be, and

Downloaded 17 Dec 2010 to 189.166.191.199. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions

239

Pu-Be, whose activities are 4.34E(8), 3.7E(9), and 1.85E(11) Bq respectively. Each
source was measured in different experimental conditions. The source-to-detector
distance for 252Cf was 120 cm, for 241Am-Be it was 50 cm, while for 239Pu-Be it was
100 cm. The count rates measured with the BSS were utilized to unfold the neutron
spectra and to calculate the dosimetric features using the ANN technology. Also, the
spectra were unfolded with the BUNKIUT code and the UTA4 matrix.

RESULTS
In Fig. 3 the neutron spectra of 252Cf, 241Am-Be, and 239Pu-Be unfolded with ANN
are shown, while the spectra unfolded with BUNKIUT code are in Fig. 4.

[ cm-2 - u-1 ]

ANN
239

Pu-Be
Am-Be
252
Cf

241

E E(E)

30
20
10
0
10-8

10-7

10-6

10-5

10-4

10-3

Neutron energy

FIGURE 3.

239

10-2

10-1

100

101

102

[ MeV ]

Pu-Be, 241Am-Be and 252Cf, neutron spectra unfolded with ANN.

By comparing the spectra in figure 3 with those in Fig. 4, it can be noticed that they
are similar. The spectra have two peaks, one is in the thermal region another is
between 0.1 to 10 MeV. Neutrons from 0.1 to 10 MeV are those produced directly by
the source; these are transported out from the source interacting with air, floor, ceiling,
walls and any media in the room, from these interactions neutrons loose energy
returning back to the room as thermal. This effect is known as room-return [9, 10].

Downloaded 17 Dec 2010 to 189.166.191.199. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions

In the epithermal region, located from 1E(-6) to 1E(-1) MeV, the spectra in Fig. 3
looks different to those in Fig. 4 because the spectra utilized to train the ANN have a
large contribution of epithermal neutrons; however, in quantitative terms the
difference is meaningless.
70

[ cm-2 - u-1 ]

BUNKIUT
239

241
252

Pu-Be
Am-Be
Cf

E(E)

10
0
10-8

10-7

10-6

10-5

10-4

10-3

Neutron energy
FIGURE 4.

239

10-2

10-1

100

101

102

[ MeV ]

Pu-Be, 241Am-Be and 252Cf, neutron spectra unfolded with BUNKIUT.

The dosimetric features obtained with the ANN for the neutron sources are in Table
1. It can be noticed that the use of ANN delivers 13 values of three types of dosimetric
magnitudes, E, H*(10), and Hp. The uncertainty of dosimetric values is 5%.
TABLE 1.
Dose [pSv-cm2]
EAP
EPA
ERLAT
ELLAT
EROT
EISO
H*(10)
Hp(10, 0o)
Hp(10, 15 o)
Hp(10, 30 o)
Hp(10, 45 o)
Hp(10, 60 o)
Hp(10, 75 o)

239

Pu-Be, 241Am-Be and 252Cf dosimetric features obtained with ANN.

239
241
252
Pu-Be
Am-Be
Cf
271
248
265
189
170
171
103
94
89
118
107
104
174
160
161
135
125
125
282
270
318
293
284
330
291
280
326
301
286
336
290
276
317
264
246
279
187
170
178

Downloaded 17 Dec 2010 to 189.166.191.199. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions

Spectra unfolded with BUNKIUT were used to fold the H*(10) resulting in 296
5%, 265 6%, and 314 4% pSv-cm2 for 239Pu-Be, 241Am-Be, and 252Cf respectively.
These values compares well with those in Table 1 for H*(10).

CONCLUSIONS
The spectra and dosimetric features of three neutron sources measured in different
experimental conditions have been obtained using the ANN technology. The input
required to do these calculations is the count rates of a BSS. Obtained results were
compared with those obtained with the BUNKIUT code, which requires more
information than the BSS count rates. Obtained results with both methods agree well.

ACKNOWLEDGMENTS
This work is part of project LINAC partially supported by the COZCyT from
Zacatecas Mexico.

REFERENCES
1. A.V. Alevra, Radioprotection 34, 305-333 (1999).
2. R. Bedogni, A. Esposito, M. Chiti and M. Angelone, Radiat. Prot. Dosim. 126, 541-545 (2007).
3. H. R. Vega-Carrillo, B. W. Wehring, K. G. Veinot and N. E. Hertel, Radiat. Prot. Dosim. 81, 133139 (1999).
4. K.A. Lowry and T.L. Johnson, Naval Research Laboratory memorandum 5340 (1984).
5. R. Bedogni, Radiat. Eff. Def. Sol. 164, 15-30 (2009).
6. S. Haykin, Neural networks, Macmillan College Publishing Co, Englewood Cliffs, NJ, 1994.
7. J.M. Ortiz-Rodriguez, M. R. Martinez-Blanco, E. Gallego, and H. R. Vega-Carrillo, Trans. Am.
Nucl. Soc., 101, 647-648 (2009).
8- J.M. Ortiz-Rodriguez, M. R. Martinez-Blanco, and H. R. Vega-Carrillo, cerma 2009, 113-118
(2009).
9 H.R. Vega-Carrillo, E. Manzanares-Acua, M. P. Iiguez, E. Gallego and A. Lorente, Radiat. Meas.,
42, 413-419 (2007).
10 H.R. Vega-Carrillo, E. Manzanares-Acua, M. P. Iiguez, E. Gallego and A. Lorente, Radiat. Meas.,
42, 1373-1379 (2007).

Downloaded 17 Dec 2010 to 189.166.191.199. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions

Neutron spectrometry and dosimetry based on a new approach called

Genetic Artificial Neural Networks
Jos Manuel Ortiz-Rodrgueza,c, Ma. del Rosario Martnez-Blancob, Eduardo
Gallegod, Hctor Ren Vega-Carrilloa,b,c*
a

Universidad Autnoma de Zacatecas, Unidad Acadmica de Ingeniera Elctrica, Av.

Ramn Lpez Velarde #801, C.P. 98000, Zacatecas, Zacatecas, Mxico.
b

Universidad Autnoma de Zacatecas, Unidad Acadmica de Estudios Nucleares, C.

Ciprs #10, Fracc. La Peuela, C.P. 98000, Zacatecas, Zacatecas, Mxico.

Departamento de Electrotecnia y Electrnica, Escuela Politcnica Superior, Avda.

Menndez Pidal, s/n - Escuela Politcnica Superior, Crdoba Spain.
d

Polytechnic University of Madrid, Nuclear Engineering Department, Jos Gutierrez

Abascal, 2, E-28006, Madrid, Spain.
Abstract. Artificial Neural Networks and Genetic Algorithms are two relatively young research areas that were
subject to a steadily growing interest during the past years. The structure of a neural network is a significant
contributing factor to its performance and the structure is generally heuristically chosen. The use of evolutionary
algorithms as search techniques has allowed different properties of neural networks to be evolved. This paper
focuses on the intersection on neural networks and evolutionary computation, namely on how evolutionary
algorithms can be used to assist neural network design and training, as a novel approach. In this research, a new
evolvable artificial neural network modelling approach is presented, which utilizes an optimization process based
on the combination of genetic algorithms and artificial neural networks, and is applied in the design of a neural
network, oriented to solve the neutron spectrometry and simultaneous dosimetry problems, using only the count
rates measured with a Bonner spheres spectrometer system as entrance data.
KEYWORDS: Neutron spectrometry, neutron dosimetry, genetic algorithms, artificial neural
networks, evolvable artificial neural networks.
1. Introduction
In recent years, the characterization of ionizing radiation fields in workplaces is one of the challenging
activities in radiation protection over the world [1]. Nowadays, neutrons are widely used in many fields
of both research and technology [2]. The workers subject to these radiations especially those who are
submitted to neutron risk have to be well monitored and protected according to relevant national
regulations which are more and more restrictive. As a result, there is an increasing demand in the field
of radiation protection to quantify these various neutron fields and to determine the radiation doses
involved [1]. The dosimetry of neutron radiation is one of the most complicated tasks in radiation
protection [3] mainly because neutron dosimetry is a complex technique [4] and highly neutron energy
dependent, and a precise knowledge on neutron spectrometry is highly essential for all dosimetryrelated studies as well as many nuclear physics experiments. In consequence, it becomes necessary to
develop additional measuring techniques to enhance the actual workers monitoring systems.
One of the suitable approaches to improve the knowledge on neutron radiation fields to which
individuals are exposed during their work, is based on spectrometric measurements [5-8]. The measured
yield of a radionuclide is the convolution of the neutron energy distribution with the response function
of the spectrometer summed over the interaction energy range. At any given angle, the measured
activity of a particular type is related to the neutron spectrum by equation (1), which represents the
discrete form of the Fredholms integral equation of the first kind without any unique solution [9].
*

Presenting author, E-mail: fermineutron@yahoo.com

C j = Ri , j * i

j = 1,2,..., m ,

(1)

i =1

where Cj is the jth detectors count rate; Ri,j is the jth detectors response to neutrons at the ith energy
interval; i is the neutron fluence within the ith energy interval; and m is the number of spheres utilized.
One of the earliest devices used for neutron spectrometry is the system known as Bonner sphere
spectrometer (BSS) system, which has been used over the years for a variety of measurements and
under various conditions, and it remains one of the most useful spectrometers [10-14]. The main
advantage of this system is its ability to cover the full range of neutron energies encountered in radiation
protection application and its isotropic response. By using the integral counts of the BSS system it is
possible to unfold the neutron spectra. The BSS response matrix, the count rates and the neutron
spectrum are related through the equation 1, which is an ill-conditioned equations system with an
infinite number of solutions. However, the weight, time consuming procedure, the need to use an
unfolding procedure and a very expert user, and the low resolution spectrum are the BSS system
drawbacks [4, 6, 9].
The current interest in neutron spectrometry has stimulated the development of several deconvolution
procedures like SAND, BUNKI, FRUIT, MAXED, GRAVEL, ANDI-03, etc. to attain improved energy
resolution through spectrum unfolding. The spectral information needs to be unfolded from the BSS
system detector responses by using any suitable computational codes such as BUNKI, MAXED, SAND,
etc., based on several methods like least square, iterative, Monte-Carlo. However, a serious drawback
arise by using these methods, which is that a lot of experience is necessary to unfold neutron spectra
using these codes, and a large number of issues can arise such as multiple stable solutions, a unique
solution with negative parts, and oscillating behaviour or spurious peaks appearing after a number of
iterations. To overcome these drawbacks, alternative approaches have been studied and proposed, to
make an efficient neutron dosimetry, and several unfolding procedures combined with various types of
experimental methods have been reported in the recent past such as Genetic algorithms (GA) [15, 16],
and Artificial Neural Networks (ANN) [9, 17-21].
Many of the previous studies in neutron spectrometry and dosimetry by using the ANN approach have
found serious drawbacks in the ANN design process itself, mainly in the proper determination of the
structural and learning parameters of the networks being designed. These parameters are significant
contributing factors to the ANN performance, however, the optimal selection of these parameters
follows in practical use no rules, and their value is at most arguable, mainly because they are generally
heuristically chosen by using the trial and error technique, which produces poor artificial neural
networks with low generalization capacity and poor performance. For the anterior, the nuclear research
community needs approaches that implement ANN models faster than what is currently available. In
consequence, more research has been suggested in order to overcome these drawbacks [ 9].
In general, the development of a good ANN model depends on several factors [22-28]. The first factor
is related to the data being used, the second factor is the network architecture or model structure, the
third factor is the model size and complexity and finally, the quality of the ANN model is also strongly
dependent on network training. This stage is essentially an identification of model parameters that fits
the given data; and is perhaps the most important factor among all. Training a neural network involves
updating the weights in such a manner that the error between the outputs of the neural networks and the
actual response being modeled is minimized.
Until today, many researchers still prefer use the gradient search method Back Propagation (BP) in
training ANN. However, this technique is a local search method and when applied to complex nonlinear
optimization problems, can sometimes result in inconsistent and unpredictable performances. One of the
main hindrances is due to the fact that searching of optimal weights is strongly dependent on initial
weights and if they are located near local minima, the algorithm would be trapped; if the initial guess of
the ANN is near local maxima, it will climb the gradient and get stuck. Several different attempts have

been proposed by various researchers to alleviate the training problems. These include imposing
constraints on the search space, restarting training at many random points, adjusting training parameters
and restructuring the ANN architecture. However, some approaches are problem-specific and not well
accepted and different researchers tend to prefer different methodologies.
Among the limitations of ANN, the followings should be given added emphasis: network architecture,
training algorithm, training data, and process relationship. All of these limitations have motivated
researchers to generate ideas of merging or hybridizing ANN with other approaches in the search for
better performance. Some of the available schemes include expert systems, statistical methods [29],
fuzzy logic, wavelet transform and as well as GA [30-33]. Aiming at improving the accuracy and
robustness of ANN model, GA was introduced in ANN modeling for connection weights or architecture
evolution [34-36]. The integration of the evolutionary computational technique into ANN models is
called Neuro-Evolutionary technique which can be used to enhance the learning capability of the ANN
model. This technique is also used to determine the suitable topology of a network and to select the
proper learning law. By evolving the ANN some of the mentioned drawbacks of conventional ANN can
be overcome.
At present, one of the more promising techniques to design the structural and learning parameters of
ANN is by introducing adaptation of network training using GA. The aim of the present work is focused
in analyzing the intersection of the artificial neural networks and the GA, analyzing like it is that GA
can be used to help in the design processes and training of ANN, i.e., in the optimum selection of the
structural and learning parameters of ANN, improving its generalization capacity, in such a way that the
neural network designed is able to unfold in an efficient way neutron spectra and to calculate equivalent
doses, starting only from the count rates obtained of a Bonner spheres spectrometric system.
2. Materials and methods

There are two major attempts to evolve ANN. The first attempt is to conjugate GA with ANN by the
evolution of connection weights. The evolution process can further divided into two stages. The first
stage only dealt with the offline set-up of the network that GA was used in training of the connection
weights instead of using classical training algorithm. The second stage extended the approach of offline training, by means of enabling on-line retraining strategy to connections weights in a fixed
topology network. This increased the capability of ANN to follow real-time interactive environment
via continuous updating the model knowledge from retraining process data. The second attempt in
providing adaptability to the neural model is to evolve its architecture. With the term architecture,
intend not only for the topological structure, but also the transfer functions of the neurons. The
evolution of architectures allows neural network to adapt their topologies to different tasks or to a non
stationary changing environment without human intervention, introducing an approach to automatic
neural network design
In this work, 300 genetically evolved neural networks were designed, trained and tested in a time of
00:01:43, selecting the best of all of them. To train the networks were used the count rates coming
from 7 Bonner spheres as input and a data set of 187 neutron spectra and 13 equivalent doses as
outputs [8, 9, 37]. In the analysis of the 187 data used in the training and testing process of the ANN
design, were used seven neurons in the input layer (corresponding to the seven Bonner Spheres used
as input variables) and 44 neurons in the output layer (corresponding 31 of them to the spectra to be
unfolded and the 13 remaining to the equivalent doses to be calculated). In the ANN design process,
was used the Root-Mean-Square Error of Prediction (RMSEP) as objective function, which is the
square root of the sum of the squared differences between the observed and predicted values for all
observations in the test set divided by the number of such observations, to estimate the prediction
error. To design the optimal ANN was used the software NeuroGenetic Optimizer (NGO), version
2.6, from BioComp Systems, Inc. [38]. The original data were automatically subdivided by the
software into training and test sets by using a random procedure. For the 187 neutron spectra and
equivalent doses samples included in the dataset, 150 samples were used as the modelling/training set,
and the remaining samples as the test set.

According NGO, the processes of GA in the context of ANN are the following. The GA process is
based on a fundamental cyclic process. This process consists of: (1) Creating an initial population of
genotypes (genetic representations of the neural networks); (2) Building neural networks
(phenotypes) based on the genotypes; (3) Training and testing the neural networks to determine how
fit they are; (4) Comparing the fitness of the networks and keeping the best; (5) Selecting those
networks in the population which are better, discarding those which arent; (6) Refilling the population
back to the defined size; (7) Pairing up the genotypes of the neural networks; (8) Mating the
genotypes by exchanging genes (features) of the networks; (9) Mutating the genotypes in some
random fashion; Then returning back to step (2) and continuing this process until some stopping
criteria is reached or manually stops the process.
Through the process described previously, the better networks survive and their features carry forward
into future generations and are combined with others to find better and better networks for the problem
considered. This genetic search capability is much more effective than random searching, as the
genetic process of recombining features vastly improves the speed of identifying highly fit networks.
The design parameters of the ANN as well as the GA, used to evolve an ANN able to unfold neutron
spectra and to calculate equivalent doses, by utilizing the NGO, were the following: The GA
parameters - generations run: 10, population size: 30, selection was performed by the top 50%
surviving, refilling of the population was done by cloning the survivors, mating was performed by
using the TailSwap technique, mutations were performed using the following technique(s): random
exchange technique at a rate of 25%.
The ANN parameters used were: the minimum network training passes for each network were 20, the
cutoff for network training passes was 50, the input neural node influence factor used was 0, the
hidden neural node influence factor used was 0, and the limit on hidden neurons was 64. The network
type was a Fast-Back Propagation neural network, the network employed 7 inputs and 1 hidden layer
with 21 Logistic 21 Tanh 21 Linear neurons (selected by NGO), and finally there were 44 output
neurons, using a linear transfer function.
3. Results

The general idea of combining GA and ANN was the following: Information about the neural network
was encoded in the genome of the genetic algorithm. At the beginning, a number of random
individuals were generated. The parameter strings were evaluated, which means a neural network was
designed according to the genome information. Its performance was determined after training with
backpropagation. Then, they were evaluated and ranked. The fitness evaluation may take more into
consideration than only the performance of the individual. Finally, crossover and mutation created
new individuals that replaced the worst - or all - members of the population. This general procedure
was quite straight-forward.
The best genetic ANN was found on generation 7 after a runtime of 00:01:43 with the following
architecture 7 inputs and 1 hidden layer with 21 Logistic 21 Tanh 21 Linear neurons and 44 output
neurons, using a linear transfer function. The Root Mean Squared Error reached on training set was
0.0191 and the Min. Root Mean Squared Error reached on test set was 0.017. Once concluded the
training and testing processes of all the designed RNA by means of NGO, it was observed that the best
RNA selected, it was even able to continue learning, the anterior makes infer that if the parameters of
the AG used to evolve the net in question, are modified, the ANN performance as well as the
generalization capability can be improved, producing in consequence better neutron spectra and
calculated doses.
The tool explored in this work, denominated Genetic Artificial Neural Networks (GANN), it is
broadly well-known for the researchers, and however, it is applied for first time in the nuclear sciences
area, specifically for the neutron spectrometry and dosimetry research. However, several drawbacks
were observed with the use of NGO applied in the solution of the neutron spectra unfolding and
equivalent dose calculus problems. The first one was that NGO has not the capability to realize
statistical tests to validate the obtained results. The second one was that NGO has not the ability to

separate the information of the unfolded neutron spectra and the 13 equivalent doses calculated, to be
presented to the end user in a proper way. Thus, the spectra and dose calculus information is presented
in a non convenient way for its interpretation to the researcher.
Because the novelty of this methodology can be observed that there are no tools to overcome the
drawbacks before mentioned and encountered during this research, being necessary to develop
scientific tools for such end. Because the anterior was designed and developed an automated tool, in
the Matlab programming environment, which is capable to realize the tasks before mentioned.
The designed GANN computer tool is capable to develop statistical tests between the desired and the
obtained data as the correlation, which is an indicative of the generalization capacity of the ANN
being designed. Must be highlighted that the correlation as an indicator of the ANN generalization
capability, has not been reported before in similar works of the literature reviewed previously, being
reported for first time in this work.
Another statistical tool included in the GANN computer tool is the Chi square test, which is an
indicator of the error encountered between the desired data and the obtained. For this work, the Chi
square test was established with a value of 18.5 for a confidence level of 95% for 30 degree of
freedom for all the neutron spectra (187) and a value of 5.226 for a confidence level of 95% for 12
degree of freedom for all the calculated equivalent doses (187).
The designed GANN computer tool also has the capability to compare, in a graphical way, the neutron
spectra unfolded and the equivalent doses, desired and obtained, just reading a spreadsheet edited for a
such end, which contain the information of the neutron spectra an equivalent doses desired and the
neutron spectra and doses obtained by means of NGO and extracted for its analysis, as is showed in
figures 1 through 4.
Figure 1, shows the best neutron spectra obtained with NGO and its corresponding 13 equivalent doses
among the 187 spectra utilized as data set, after the training and testing stages. In this figure can be
seen that, in this case, the neutron spectra is better than the equivalent doses calculated. However
according the statistical test effectuated by means of the GANN tool, showed in figure 1, can be seen
that there is not statistical difference between the neutron spectra unfolded and the 13 equivalent doses
desired (red line) and obtained (blue line), because the value obtained for the Chi square test for both
neutron spectra and doses, it is so inferior to the critical value of the Chi square test (18.5). At the
same time, the value obtained applying the correlation test, near at optimum value equal to one,
reveals that the ANN generalization ability is high.
Figure 1: Best Neutron spectra and the corresponding 13 equivalent doses obtained with NGO.

Figure 2, shows the better equivalent doses calculated and its corresponding neutron spectra obtained
with NGO after training and testing of the better ANN. In this figure can be observed that opposite to
figure 1, the equivalent doses are better than the neutron spectra unfolded, however the statistical tests
show that there are not differences between the spectra and doses desired (blue line) and obtained (red
line). In a similar way the generalization capacity of the designed ANN is high according the
correlation test effectuated for both neutron spectra and doses.
Figure 2: Best equivalent doses calculated and the corresponding neutron spectra obtained with NGO

Figure 3, shows the worst neutron spectra and its corresponding 13 equivalent doses calculated (red
line) compared with the desired data (blue line). Newly in figure 3, can be seen that after applying the
statistical test there are not statistical differences between the desired data and the obtained. In this
case the generalization capacity is low in the neutron spectra case and high in the calculated doses case
according the results obtained after applying the correlation test, as can be appreciated in figure 3.
Figure 3: Worst neutron spectra and the corresponding 13 equivalent doses obtained with NGO

Figure 4, shows the worst 13 equivalent doses and its corresponding neutron spectra. In a similar way
as in the case 3, after applying the statistical tests, was observed that there are not statistical
differences among the desired (red line) and obtained (blue line) data. From figures 3 and 4 can be
seen that although the reported cases are the worst, the generalization capability of the ANN designed
by means of NGO is high for the calculated doses and low for the neutron spectra. This pattern is
repeated for the majority of the cases effectuated (187 neutron spectra and equivalent doses). 140 from
the 187 cases have a correlation value equal or bigger than 0.9 for the equivalent doses and 28 from
the 187 cases have a correlation value equal or bigger than 0.9 for the neutron spectra. The anterior

reveals that the ANN designed by means of NGO, is capable to train better for the equivalent doses
case. According the results obtained in this work, it is evident that more research is needed, in order to
improve the ANN performance, because the combination of GA and ANN seems to be a very useful
approach.
Figure 4: Worst equivalent doses and the corresponding neutron spectra obtained with NGO

4. Conclusions

In this research, a new evolvable ANN modelling approach called Genetic Artificial Neural networks
(GANN) methodology was presented, which utilizes an optimization process based on the
combination of genetic algorithms and artificial neural networks applied in the ANN design, oriented
to solve the neutron spectrometry and simultaneous dosimetry problems, using only the count rates
measured with a Bonner spheres spectrometer system as input. The experiments described in the
present work, demonstrate that the GA technology can be applied for finding the optimal ANN
parameters and architecture in the neutron spectrometry and dosimetry domains of nuclear sciences,
reducing significantly the time spent in the ANN design process itself if compared with the trial and
error approach, reported in prior works.
The correlation test as indicator of the generalization capability of the ANN being designed, it is
reported for first time in this work, and it is a useful tool to analyze the performance of the network in
early stages of the training and testing of the ANN.
During the ANN design process, by means of NGO, was observed that the best ANN designed,
showed a tendency to train better the information corresponding to the equivalent doses, which is
evident when analyzing the correlation and Chi Square tests of the data set desired and obtained with
NGO by means of the GANN computer tool.
Once concluded the training and testing processes of all the designed RNA by means of NGO, it was
observed that the best RNA selected, it was even able to continue learning, the anterior makes infer
that if the parameters of the AG used to evolve the net in question, are modified, the ANN
performance as well as the generalization capability can be improved, producing in consequence better
neutron spectra and calculated doses. However, a new problem arise which is the proper selection of
the GA parameters for the ANN design. At present, work is being done to this respect in order to solve
these serious drawbacks.
The genetic search capability in ANN design is much more effective than random searching
approaches, as the genetic process of recombining features vastly improves the speed of identifying
highly fit networks. It also has a potential advantage over just using personal experience in building
neural networks, as new and potentially better solutions may be found through this process than might

be found using the nearly unavoidable assumptions made by the user. The new ideas and concepts of
GA and ANN bring new life into Artificial Intelligence research. However, the problem of combining
GA and ANN lies in the encoding of the network.
Unlike BP training algorithm of ANN, GA is a global search algorithm based on the principle
survival of fittest. It simultaneously searches for solutions in several regions, thus increasing the
probability of global convergence. Furthermore, since it is impossible to formulate an a priori exact
model of the system, a more practical approach is off-line set up a rough model, followed by on-line
update of the model using GA. In this way, the merging of GA and ANN will gain adaptability to
dynamic environment and lead to significantly better intelligent systems than relying on ANN or GA
alone. In other words, the GA- ANN model presented here should be more robust to dynamic
nonlinearity of the process involved.

Acknowledgments

This work is part of the SYNAPSIS project supported by CONACyT (Mxico) under contract SEP2004-C01-46893.

REFERENCES

[1]

[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]

MAZROU, H., et al., Characterization of the CRNA Bonner sphere spectrometer based on 6LiI
scintillator exposed to an 241AmBe neutron source, Radiation Measurements, 43, (2008) 10951099.
WIELUNSKI, M., Wet al., Intercomparison exercise with MeV neutrons using various
electronic personal dosimeters, Radiation Measurements, 43, (2008) 1063-1067.
SOUTO, E.B. et al., Development of a brazilian-neutron dosimeter, Nuclear Instruments and
Methods in Physics Research, 226 (B), (2008) 3174-3177.
BEGDONI, R., et al., FRUIT: An operational tool for multisphere neutron spectrometry in
workplaces, Nuclear Instruments and Methods in Physics Research, A 580, 1301-1309 (2007).
Thomas J. D., Neutron spectrometry for radiation protection, Radiation Protection Dosimetry,
110 (2004).
BROOKS, F.D. et al., Neutron spectrometry historical review and present status, Nuclear
Instruments and Methods in Physics Research, 476 (A), (2002) 1-11.
McDONALD, J.C., et al., Neutron spectrometry for radiation protection purposes, Nuclear
Instruments and Methods in Physics Research, A, 476 (2002).
VEGA-CARRILLO, H.R, et al., Neutron spectra re-binning and dose calculation using Monte
Carlo methods, Revista Mexicana de Fsica, 53, (2007) 3.
VEGA-CARRILLO, H.R, et al., Neutron spectrometry using artificial neural networks,
Radiation Measurements, 41, (2006) 425.
MATZKE M., Unfolding procedures, Radiation Protection Dosimetry, 107, (2003).
EL MESSAOUDI, M., et al., Performance of three different unfolding procedures connected to
Bonner sphere data, Radiation Protection Dosimetry, 108, (2004).
LACOSTE V., et al., Characterization of Bonner spheres systems at monoergetic and thermal
neutron fields, Radiation Protection Dosimetry, 110, (2004).
VYLET, V., Response matrix of an extended Bonner sphere system, Nuclear Instruments and
Methods in Physics Research, A, 476, (2002).
VEGA-CARRILLO, H.R, et al., Response matrix of a multisphere neutron spectrometer with
an 3He proportional counter, Revista Mexicana de Fsica, 51, (2005).
MUKHERJEE, B., A high-resolution neutron spectra unfolding method using the genetic
algorithm technique, Nuclear Instruments and Methods in Physics Research, A, 476, (2002).
FREEMAN, D.W., et al., Genetic algorithms-A new technique for solving the neutron
spectrum unfolding problem, Nuclear Instruments and Methods in Physics Research, A, 425,
(1999).

[17] VEGA-CARRILLO, H.R, et al., Artificial neural networks in neutron dosimetry, Radiation
Protection Dosimetry, 118, (2006) 251.
[18] BRAGA, C.C. et al., Application of neural networks for unfolding neutron spectra neasured by
means of Bonner spheres, Nuclear Instruments and Methods in Physics Research, A, 476
(2002).
[19] CORDES, A., Fehrenbacher et al., An approach to unfold the response of al multi-element
system using artificial neural networks, Wahl W., IEEE Transaction of Nuclear Science, 45
(1998).
[20] FEHRENBACHER, G., et al., Proposal of a new method for neutron dosimetry based on
spectral information obtained by application of artificial neural networks, Radiation Protection
Dosimetry, 83, (1999).
[21] KARDAN, M.R., et al., Fast neutron spectra determination by threshold activation
detectorsusing neural networks, Radiation Measurements, 38, (2004).
[22] FAUSETT, L., Fundamentals of neural networks, Prentice-Hall, (1993).
[23] GUPTA, M.M., et al., Static and Dynamic Neural Networks. From Fundamentals to Advanced
Theory, John Wiley & Sons, (2003).
[24] HAYKIN, S., Neural Networks: A Comprehensive Foundation, Prentice Hall, (1999).
[25] JAIN, A.K., et al., Artificial neural networks: a tutorial, IEEE, 29, (1996).
[26] LIPPMANN, R.P., An introduction to computing with neural networks, IEEE ASSP
MAGAZINE, 4, (1987) 2.
[27] GUPTA, M.M., et al., Static and Dynamic Neural Networks from Fundamentals to Advanced
Theory, John Wiley & Sons, (2003).
[28] KOHONEN, T., The self-organizing map, Procceedings IEEE, 78, (9), (1990) 1464-

1480.
[29] ORTIZ-RODRGUEZ, J.M., et al., Robust design of artificial neural networks applying the
Taguchi methodology and DoE, CERMA (2006).
[30] GOLDBERG, David E., Genetic Algorithms in Search, Optimization, and Machine Learning,
Addison-Wesley (1989).
[31] RANDY, L. Haupt and Sue Ellen Haupt, Practical genetic algorithms, Wiley (2004).
[32] FRANZ, R., Representations for Genetic and Evolutionary Algorithms, Springer (2006).
[33] MITCHELL, M., An Introduction to Genetic Algorithms, MIT Press (1998).
[34] XIN, Yao, A review of evolutionary artificial neural networks, International Journal of
Intelligent Systems, 8, (1993).
[35] LAKHMI, C., et al., Fusion of Neural Networks, Fuzzy Systems and Genetic Algorithms:
Industrial Applications, CRC Press (1998).
[36] WHITESON, S., et al., Evolutionary Function Approximation for Reinforcement Learning,
Journal of Machine Learning Research, 7, (2006).
[37] VEGA-CARRILLO, H.R, et al., Catalogue to select the initial guess spectrum during
unfolding, Nuclear Instruments and Methods in Physics Research, A, (2002).
[38] BIOCOM SYSTEMS, (2008), http://www.biocompsystems.com

A Neutron spectrometry and dosimetry computer tool based on ANN

Jos Manuel Ortiz-Rodrgueza,c, Ma. del Rosario Martnez-Blancob, Eduardo
Gallegod, Hctor Ren Vega-Carrilloa,b,c*
a

Universidad Autnoma de Zacatecas, Unidad Acadmica de Ingeniera Elctrica, Av.

Ramn Lpez Velarde #801, C.P. 98000, Zacatecas, Zacatecas, Mxico.
b

Universidad Autnoma de Zacatecas, Unidad Acadmica de Estudios Nucleares, C.

Ciprs #10, Fracc. La Peuela, C.P. 98000, Zacatecas, Zacatecas, Mxico.

Departamento de Electrotecnia y Electrnica, Escuela Politcnica Superior, Avda.

Menndez Pidal, s/n - Escuela Politcnica Superior, Crdoba Spain.
d

Polytechnic University of Madrid, Nuclear Engineering Department, Jos Gutierrez

Abascal, 2, E-28006, Madrid, Spain.
Abstract. In the neutron spectrometry and dosimetry research areas by means of the Bonner spheres
spectrometric system utilizing classical approaches, such as Monte Carlo, parametrization and iterative
procedures, the weight, time consuming procedure, the need to use an unfolding procedure, the low resolution
spectrum, and the need to use the neutron fluence-to-dose conversion coefficients for calculating the equivalent
doses are some drawbacks which these approaches offer. Each of the mentioned difficulties has motivated the
development of complementary procedures such as maximum entropy, genetic algorithms and artificial neural
networks. The use of neural networks to unfold neutron spectra and to calculate equivalent doses from the count
rates measured with BSS system has become in an alternative procedure, which has been applied with success,
however, it is observed that doesn't exist computer tools based on ANN technology to unfold neutron spectra and
to calculate equivalent doses. In this work a customized front end user interface software application, called
Neutron Spectrometry and Dosimetry by means of Artificial Neural Networks (NSDANN) based on ANN
technology, is presented, which is capable to unfold neutron spectra and to simultaneously calculate 13
equivalent doses, by using only the count rates of a BBS system as input, in just a few seconds.

KEYWORDS: Artificial Neural Netowork, Neutron Spectrometry, Neutron Dosimetry, Software

Development.
1. Introduction
The dosimetry of neutron radiation is one of the most complicated tasks in radiation protection [1]. This
arises from several factors, related to either the definition of the operational quantities, or the practical
problems in the instrument design and calibration [2]. Reliable determination of neutron doses in
workplaces is still an issue in the field of radiation protection [3].
Although there is a wide range of different devices used for neutron spectrometry, the majority of the
instruments can be grouped together into a small number of broad categories, each one based on a
common underlying technique [4-6]. Among the many available neutron spectrometry techniques, the
multisphere or Bonner sphere spectrometer (BSS) system is the most used for radiation protection
purposes [7-9], due to advantageous characteristics as wide energy range (from thermal to GeV
neutrons), large variety of active or passive thermal sensors allowing adapting the sensitivity to the
specific workplace, good photon discrimination and simple signal management. Disadvantages are the
poor energy resolution, which does not allow appreciating fine structures as narrow peaks, the weight,
and the need to sequentially irradiate the spheres, requiring, in general, long exposure periods [1].
In the neutron spectrometry by means of the BSS system and utilizing the classical approaches, as for
example Monte Carlo, parametrization and iterative procedures, the weight, time consuming procedure,
*

Presenting author, E-mail: fermineutron@yahoo.com

the need to use an unfolding procedure, the low resolution spectrum, and the need to use the neutron
fluence-to-dose conversion coefficients for calculating the equivalent doses are some drawbacks which
these approaches offer. Each of the mentioned difficulties has motivated the development of
complementary procedures such as maximum entropy, Genetic Algorithms (GA) and Artificial Neural
Networks (ANN). The use of ANN to unfold neutron spectra and to calculate equivalent doses from the
count rates measured with BSS system has become in an alternative procedure [10-15], which has been
applied with success, and some authors have expressed that this technology merit attention because de
success obtained [2].
In the ANN design process, the choice of the ANNs basic parameters (network topology, learning rate,
initial weights, etc) often determines the success of the training process. The selection of these
parameters follows in practical use no rules, and their value is at most arguable, because users have to
choose the architecture and determine many of the parameters in a selected network by employing the
trial and error technique which produces ANN with poor performance and low generalization capability
spending often large amount of time. To overcome these drawbacks a new approach known as Robust
Design of Artificial Neural Networks (RDANN) methodology has been proposed [22-27]. The RDANN
methodology is a powerful method based on parallel processes where all the experiments are planned a
priori and the results are analyzed after all experiments are completed. RDANN can help to provide
answers to critical design and construction issues of ANN.
As mentioned before, by using the integral counts of the BSS system it is possible to unfold the neutron
spectra by using several methods like Monte Carlo, parametrization and iterative procedures, considered
as classical approaches. The use of ANN to unfold neutron spectra and to calculate equivalent doses
from the count rates measured with BSS system has become in an alternative procedure which has been
applied with success [10-15], however, the optimum ANN parameter determination problem remains.
RDANN has been used to alleviate this drawback [22-27].
Provided a well-established response matrix and adequate irradiation conditions, the most delicate part
of the BSS-based spectrometry are the unfolding process. Many computerized BSS unfolding codes
have been developed, as for instance BONDI-97, SAND, MAXED, BUNKIT [28]. Other than
conventional methods used to unfold the neutron spectra, considered as new approaches, have recently
been developed such as MITOM [29], FRUIT [2], GRAVEL [3,30] and NUBAY [3,30]. The critical
points of these neutron spectra unfolding codes are, in general, the complexity of the codes, the need of
a very expert user and the need of realistic a priori information, such as a default spectrum as close
as possible to the spectrum to be obtained [2]. These considerations, together with those suggested by
the experience in radiation protection around neutron producing facilities, suggest that an unfolding
code devoted to the operational workplace neutron monitoring would be of great help to the radiation
protection community [2]. With this purpose, a new computer tool based on ANN technology was
developed. In this work is presented a customized front end user interface software application, called
Neutron Spectrometry and Dosimetry by means of Artificial Neural Networks (NSDANN) based on
ANN technology, capable to unfold neutron spectra and to simultaneously calculate 13 equivalent
doses, by using only as input data the count rates coming of a BBS system, in just a few seconds if
compared with the time spent with the classical techniques, not being needed a priori information about
the spectra being calculated.
2. Materials and methods
In prior works, a large variety of neural network architectures and topologies were developed, by using
the RDANN methodology, in order to build an ANN capable to solve the neutron spectrometry and
dosimetry problems. The best ANN selected was: network topology: 7:26.39:44, momentum: 0.1,
learning rate: 0.1, training algorithm: trainscg, target mean square error: 1e-4, as is showed in figure 1,
where can be seen a basic scheme of the ANN designed. In this figure, can be highlighted that the ANN
requires only as input data, the count rates coming from a BBS system; the neutron spectra is unfolded
and 13 equivalent doses are simultaneously calculated in an automated way, after just a few seconds and
in an independent way, i.e., for calculating the equivalent doses, this ANN does not require the
spectrometric information.

Figure 1: Neutron spectrometry-dosimetry based on ANN

Once the networks parameters were determined (which was the harder task), the same was trained and
tested, however, from the literature reviewed, it was observed that doesn't exists computer tools based
on ANN technology to make neutron dosimetry in nuclear workplaces, in such a way that it exists the
necessity for building a computational tool for such an end, which is the aim of this work.
The NSDANN computer tool was designed under Matlab programming environment extracting the
numerical data stored at synaptic weights, and building an end user interface as is showed in figure 2,
where can be seen that the NSDANN computer tool requires as entrance data, only the measured
Bonner spheres counts rate which must be stored in an spreadsheet in a local folder of the computer.
After the counts rate are reading by NSDANN, it produces the output, i.e., the neutron spectra unfolded
and 13 equivalent doses, after a few seconds in a completely automated way. The information produced
by NDSANN is available in both graphical and numerical format for a further analysis and
manipulation of the researcher.
Figure 2: NSDANN computer tool designed with ANN

By using NSDANN tool, showed in figure 2, two cases were carried out in order to probe and to
validate the designed tool comparing the results obtained against classical and contemporary neutron
spectrometry approaches. The first one case was as follows: a BBS system with a 0.4x0.4 cm2 6LiI(Eu)
scintillator as thermal neutron was used to obtain the counts rate of a of 252Cf neutron source at 100 on
air, in the Unidad Acadmica de Estudios Nucleares of the Universidad Autnoma de Zacatecas (UAZ),
Mexico. Then the obtained data were introduced to NSDANN. The results were compared against the
results obtained for the same count rate but using the neutron spectra unfolding code known as BUNKI.
For the second case, was used a data set obtained from an experiment reported and carried out with a
new neutron spectrometric tool (MITOM) proposed by Universidad Autnoma de Barcelon (UAB),
Spain [31]. Several important differences exist between the UAB experiment by means of MITOM and
the UAZ experiment by means of NSDANN tool.

The most important difference relies in the type of detector used in each case. In figure 3 can be seen the
response matrix of the UAB for 12 Bonner spheres and a detector of the type 3He, whereas in figure 4
can be seen the response matrix of the UAZ for 7 Bonner spheres and a cm2 6LiI(Eu) scintillator as
detector. It is important to mention that the UAZ response matrix was used to train the ANN contained
in NSDANN [27]. Table 1, shows the counts rate readings of the UAB BSS system. In table 1, can be
seen highlighted the data used to compare and validate the NSDANN computer tool.

Figure 3: BBS response matrix of the 3He detector (UAB)

Figure 4: BBS response matrix of the 6LiI(Eu) detector (UAZ)

101
0"

100

3"
5"
6"

-1

[ cm ]

8"
10"
10-2

12"

10-3

10-4

10- 10

10-9

10-8

10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

101

102

Neutron energy [ MeV ]

Table 1: Bonner spheres counts rate of the UAB BSS system for a realistic spectra used in UAZNSDANN
Sphere diameter
[inches]
2.5
3
4.2
5
6
8
10
12
2.5/Cd
3/Cd
4.2/Cd

Measured count rate

[cps]
5703e-4
6120e-4
6111e-4
5436e-4
4452e-4
2389e-4
1132e-4
5120e-4
2638e-4
3396e-4
4235e-4

3. Results
In this work was presented a front end computer tool, based on the ANN technology and RDANN
methodology, called NDSANN, with the following characteristics: it is a front end computer tool
application, simple for the use of the end user, fast in the neutron spectra unfolding process. This tool
is capable to unfold neutron spectra and simultaneously calculates 13 different equivalent doses in an
independent way, with just the count rates of the BSS system as entrance data, not requiring the
neutron spectra information to calculate the equivalent doses. The NSDANN tool after analyzing the
input data, presents the results obtained in both graphical and numerical way. From the two cases
considered in this work were obtained the following results:
Figure 5 shows the neutron spectra obtained with NSDANN (blue line) which is compared with the
obtained by means of BUNKI code (red line). Whereas Figure 6 shows 13 equivalent doses calculated
by means of NSDANN. The equivalent doses are: EAP, EPA, ERLAT, ELLAT, EROT, EISO, H*(10),
Hp,slab(10,0), Hp,slab(10,15), Hp,slab(10,30), Hp,slab(10,45), Hp,slab(10,60), Hp,slab(10,75). [14]
Figure 5: Case one - Comparison between the neutron spectra unfolded by means of BUNKI code and
NSDANN at UAZ

On the other hand, figure 7 shows the neutron spectra obtained by means of MITOM code at a nuclear
reactor in canel/T400 of Cadarache France, which is compared with the neutron spectra obtained by
NSDANN, showed in figure 8. Finally in figure 9 is shown the 13 equivalent doses calculated by
NSDANN. From the results obtained in this work, it is demonstrated that the designed ANN
(contained in the NSDANN computer tool) has a high generalization capability, because it is capable
to unfold the neutron spectra, even in the case two, where the spectrometric conditions (neutron
detector, Bonner spheres and response matrix) are too different of the trained ANN contained in the
designed NSDANN computer tool.
Figure 6: Case one 13 equivalent doses calculated by means of NSDANN (UAZ).

Figure 7: Neutron spectra unfolded by mean of MITOM tool of UAB for the reference realistic
spectra at Canel/T400 de Cadarache.

Figure 8: Neutron spectra unfolded by mean of UAZ-SDANN for the reference realistic spectra of
case two.

Figure 9: 13 equivalent doses calculated by means of NSDANN (UAZ) for case two.

From the results obtained in this research, it is evident that Matlab is a powerful programming
software for ANN design; however it presents some drawbacks which make undesirable the use of this
programming language in the implementation of the NSDANN computer tool, because of several
inconveniences such as the NDSANN tool depends totally on Matlab to be executed, if a user does not

have installed Matlab NSDANN can not be executed. Another inconvenience is that the front end
application is based in text which makes it not friendly, and although within Matlab it is possible to
develop graphical interfaces, the first drawback continues being present. Fortunately Matlab allows
extract the knowledge of the ANN trained and tested, which is stored at the synaptic weights, as
numerical data in a matrix form, which makes possible the design of a customized graphical user
interface outside of the Matlab environment. Work is being done to this respect.
4. Conclusions
In this work, a new computer tool called NSDANN, based on ANN technology, was designed to
unfold neutron spectra and to simultaneously calculate 13 equivalent doses within Matlab
programming environment. Important characteristics of NSDANN are the following: first, the user
does not need to be a very expert user; second, the tool is not complex to use for the end user, third, it
is not required a priori information about the spectra being calculated. The proposed computer tool
reduces significantly the time required to prepare, to process and to present the information, in an
appropriate way to the researcher, if compared with the traditional approaches. This tool overcome
some drawbacks of the classical methodologies, reducing the time spent in the entire neutron spectra
unfolding and dose calculus processes.
One notable characteristic of NSDANN is that for the equivalent doses calculus is not required the
spectrometric information, because in the ANN training the neutron spectra and the calculus of the
equivalent doses were considered as independent problems but they were solved simultaneously just
with one ANN.
From the examples carried out in this work can be seen that the ANN designed by means of RDANN
methodology and implemented in the NSDANN computer tool has high performance and
generalization capability, because was capable to generalize in order to solve a neutron spectra
problem with conditions very different in which the ANN was trained.
Matlab, is a powerful programming software which has, so far been successfully used for solve ANN
problems in engineering and scientific environments; however it presents some drawbacks which
make undesirable the use of this programming language because of several inconveniences. The
knowledge extraction of the synaptic weights as numerical data in a matrix way, makes possible the
design of a customized graphical user interface outside of the Matlab environment. Work is being
done to this respect.
Just two experiments were carried out with the NSDANN computer tool based on ANN technology, to
verify its effectiveness. More experiments are required in order to verify the results obtained
comparing them with another probed approaches.
The use of ANN technology is a useful alternative to solve the neutron spectrometry and dosimetry
problems; however, some drawbacks must be solved in the ANN design process, such as the proper
synaptic weight initialization, and the optimum ANN topology selection, in order to obtain the best
results which this approach is capable to offer. At present research and development of tools in this
sense are been carried out.
In future works, would be necessary to perform statistical tests to validate if exists statistical difference
between the desired data and the obtained. Work is being carried out in this sense [32]

Acknowledgments
This work is part of the SYNAPSIS project supported by CONACyT (Mxico) under contract SEP2004-C01-46893.

REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]

[11]
[12]

[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]

[23]

[24]

BEGDONI R., Neutron spectrometry and dosimetry for radiation protection around a high
energy electrn-positron collider, PhD Thesis, Universidad Autonoma de Barcelona, (2006).
BEGDONI, R., et al., FRUIT: An operational tool for multisphere neutron spectrometry in
workplaces, Nuclear Instruments and Methods in Physics Research, A 580, 1301-1309 (2007).
LACOSTE, V., et al., Bonner sphere neutron spectrometry at nuclear workplaces in the
framework of the evidos project, Radiation Protection Dosimetry, 125 (1-4), 304-208, (2007).
McDONALD, J.C., et al., Neutron spectrometry for radiation protection purposes, Nuclear
Instruments and Methods in Physics Research, A, 476, (2002).
THOMAS, J. D., Neutron spectrometry for radiation protection, Radiation Protection
Dosimetry, 110, (2004).
MATZKE, M., Unfolding Procedures, Radiation Protedction Dosimetry, 107, (2003).
EL MESSAOUDI, M, et al., Performance of three different unfolding procedures connected to
Bonner sphere data, Radiation Protection Dosimetry, 108, (2004).
LACOSTE, V., et al., Characterization of Bonner spheres systems at monoergetic and thermal
neutron fields, Radiation Protection Dosimetry, 110, (2004).
VYLET, V., Response matrix of an extended Bonner sphere system, Nuclear Instruments and
Methods in Physics Research, A, 476, (2002).
BRAGA, C.C., et al., Application of neural networks for unfolding neutron spectra measured
by means of Bonner spheres, Nuclear Instruments and Methods in Physics Research, A, Vol.
476, (2002).
CORDES, A., et al., An approach to unfold the response of a multi-element system using an
artificial neural network, IEEE Transaction of Nuclear Science, 45, (1998).
FEHRENBACHER, G., et al., Proposal of a new method for neutron dosimetry based on
spectral information obtained by application of artificial neural networks, Radiation Protection
Dosimetry, 83, (1999).
KARDAN, M.R., et al., Fast neutron spectra determination by threshold activation detectors
using neural network, Radiation Measurements, 38, (2004).
VEGA-CARRILLO, H.R., et al, Artificial neural networks in neutron dosimetry, Radiation
Protection Dosimetry, 118, (2006) 251.
VEGA-CARRILLO, H.R., et al., Neutron spectrometry using artificial neural networks,
Radiation Measurements, 41, (2006) 425.
FAUSETT, L., Fundamentals of neural networks, Prentice-Hall, (1993).
GUPTA, M.M., et al., Static and Dynamic Neural Networks. From Fundamentals to Advanced
Theory, John Wiley & Sons, (2003).
HAYKIN, S., Neural Networks: A Comprehensive Foundation, Prentice Hall, (1999).
JAIN, A.K., et al., Artificial neural networks, a tutorial, computer, IEEE, 29, (1996).
LIPPMANN, R.P., An introduction to computing with neural networks, IEEE ASSP
MAGAZINE, 4, (1987) 2.
KOHONEN, T., The self-organizing map, Procceedings IEEE, 78, (1990) 9.
ORTIZ-RODRGUEZ, J.M., et al., Problemas crticos en el diseo y construccin de redes
neuronales en la espectrometra de neutrones, utilizando la metodologa de diseo robusto de
redes neuronales artificiales, 1er. Congreso Panamericano del IRPA, VII Congreso Regional de
Seguridad Radiolgica y Nuclear, XIV Reunion Anual de la SMSR y XVII Congreso Anual de
la SNM, Acapulco, Guerrero, Mxico (2006); Contributed Papers (CD-ROM).
MARTNEZ-BLANCO M.R., et al., Neutron spectrometry and dosimetry by means of Bonner
spheres system and artificial neural networks applying robust design of artificial neural
networks, 1er. Congreso Panamericano del IRPA, VII Congreso Regional de Seguridad
Radiolgica y Nuclear, XIV Reunin Anual de la SMSR y XVII Congreso Anual de la SNM,
Acapulco, Guerrero, Mxico (2006); Contributed Papers (CD-ROM).
ORTIZ-RODRGUEZ, J.M., et al., A proposal to order the neutron data set in neutron
spectrometry using the RDANN methodology, IX International Symposium, XIX National
Congress on Solid State Dosimetry, Mexico city (2006); Contributed Papers (CD-ROM).

[25] ORTIZ-RODRGUEZ, J.M., et al., RDANN a new methodology to solve the neutron spectra
unfolding problem, IX International symposium, XIX National Congress on Solid State
Dosimetry, Mexico city (2006); Contributed Papers (CD-ROM).
[26] MARTNEZ BLANCO, M.R., et al., Neutron spectrometry and dosimetry by means of Bonner
spheres system and artificial neural networks applying robust design of artificial neural
networks, IX International Symposium, XIX National Congress on Solid State Dosimetry,
Mexico city (2006); Contributed Papers (CD-ROM).
[27] ORTIZ RODRGUEZ, J.M., et al., Robust design of artificial neural networks applying the
taguchi methodology and DoE, CERMA (2006).
[28] MILLER, S.C., AFITBUNKI: A Modified Iterative Code to Unfold Neutron Spectra from
Bonner Sphere Detector Data, Master's thesis, Air force inst of tech wright-pattersonafb oh
school of engineering, (1993).
[29] TOMS, M., et al., MITOM: A new unfolding code based on a spectra model method applied
to neutron spectrometry, Radiation Protection Dosimetry, 110 (1-4), (2004) 545-548.
[30] COECK, M., et al., Neutron spectrometry around the venus reactor using Monte Carlo
simulations and Bonner spheres measurements, Radiation Protection Dosimetry, advance
access, (2007).
[31] FERNNDEZ, F., Espectrometra aplicada a la dosimetra, Radioproteccin, 37, IX, (2002).
[32] ORTIZ RODRGUEZ, J.M., et al., Genetic Artificial Neural Networks in Neutron
Spectrometry and Dosimetry, work being accepted for its presentation at the 12th International
congress of IRPA, Buenos Aires, Argentina, (2008).

Evolutionary Artificial Neural Networks in Neutron Dosimetry

Ortiz-Rodriguez J.M1,3., Martinez-Blanco M.R.2, Gallego E.4 and Vega-Carrillo H.R1,2,3.
Universidad Autnoma de Zacatecas, U. Acadmicas: 1-Ingeniera Elctrica y 2-Estudios Nucleares
Av. R.L. Velarde No. 801, Col. Centro, C.P. 98000, Zacatecas, Mxico
3-Universidad de Crdoba en Espaa, Depto. de Electrotecnia y Electrnica, Escuela Politecnica Superior
Avda. Menendez Pidal, s/n - Escuela Politecnica Superior, Crdoba, Espaa
4-Universidad Politecnica de Madrid, Depto. de Ingeniera Nuclear
Jos Gutierrez Abascal, 2, E-28006, Madrid, Espaa
morvymm@yahoo.com.mx
INTRODUCTION

DESCRIPTION OF THE ACTUAL WORK

The dosimetry of neutron radiation is one of the most

complicated tasks in radiation protection mainly because
it is a complex technique, highly dependent of the neutron
energy, and a precise knowledge on neutron spectrometry
is highly essential for all dosimetry-related studies as well
as many nuclear physics experiments. In consequence, it
becomes necessary to develop additional measuring
techniques to enhance the actual workers monitoring
system. In the last decade, several approaches based on
Artificial Intelligence (AI) techniques, such as Genetic
Algorithms (GA) and Artificial Neural Networks (ANN)
have been reported in the literature as alternative
approaches for the neutron spectrometry and dosimetry
problems [1], starting from the count rates of the Bonner
spheres system. ANN have received the higher attention
among researchers, however, the use of this technology is
not free of problems, as the researches advance new
questions arise and new scientific and technological
problems are presented [2].
Many of the previous studies in the neutron
spectrometry and dosimetry domain by means of ANN
have found serious drawbacks in the design process of the
nets [2]. One of the most promising techniques for the
ANN design, is to introduce the capacity of adaptation to
the net by using Evolutionary Algorithms (EA), which
can be used to adapt the connections among the synaptic
weights, the design of the architecture, the adaptation of
the learning rule, etc. [3] The use of EA for the modeling
of ANN applied in the neutron spectrometry and
dosimetry problems, is an alternative technique that has
demonstrated to be convenient in the design process of
neural networks, opposite to the problems experienced by
other researchers which use the trial and error technique.
Because this is an emergent technology, at present
there are no tools for the analysis of the results obtained,
therefore, in this work is presented the design of a
computational tool known as Neutron Spectrometry and
Dosimetry Toolbox, for the analysis of the spectrometric
and dosimetric information obtained by means of the
Evolutionary Artificial Neural networks (EANN)
methodology.

In previous works, 300 EANN were designed, trained and

tested, each in a time of 0.34 seconds average. For each
one of the ANN designed, were used seven neurons in the
input layer (corresponding to seven Bonner Spheres, as
input variables) and 44 neurons in the output layer
(corresponding 31 of them to the spectra to be unfolded
and the remaining 13, to equivalent doses calculated) [4].
For the training process of the networks, a data set of 187
neutrons spectra was used, and to design the optimal
ANN evolved by means of GA the software
NeuroGenetic Optimizer (NGO). Because the novelty of
this methodology and the lack of tools for the analysis of
the spectrometric and dosimetric results obtained with
NGO, a computer code denominated Neutron
Spectrometry and Dosimetry Toolbox (NSDTB) was
designed under the LabVIEW programming environment
in a graphical user interface, oriented to the end user, in
order to overcome these drawbacks.
RESULTS
Figure 1, shows the main window of NSDTB, where
can be seen that it is divided in two sections: The upper
section shows the graphical (left side) and numerical
(right side) information of the neutron spectra unfolded.
In a similar way, the lower section shows information of
the equivalent doses calculated by the EANN. As
additional element NSDTB has the ability to realize the
Chi square statistical test among the spectra and doses
being analyzed.
The general operation of NSDTB is the following:
after executing the main program, will appear the main
window showed in figure 1. To analyze the spectrometric
and dosimetric information obtained with NGO, the user
must click in the Load data button, a window will appear
requesting to the user a text file containing the
corresponding information to analyze. The text file should
be edited previously for the user as follows: due to the
nature of the ANN training process, the first 31 rows of
each column will correspond to the spectra to be analyzed
and the remaining 13 to the equivalent doses (each
column must be separate by tabulations).

ACKNOWLEDGMENTS
This work is part of the SYNAPSIS project supported
by CONACyT (Mxico) under contract SEP2004-01-46893.
REFERENCES
1.

2.
Fig. 1. Neutron spectra and 13 equivalent doses obtained
with NGO, compared with its expected values, by means
of NSDTB.
Figure 1, shows the spectrometric and dosimetric
comparison obtained with the best EANN designed with
NGO in just 0.34 seconds, and the spectra and doses
expected contained in the data set used for the EANN
training. As can be seen in this figure, the chi square
statistical tests calculated, demonstrate that there are not
statistical differences between both spectrum and doses
expected and obtained, which make the EANN
technology a very convenient approach in the neutron
spectrometry and dosimetry domain because the time
spent in the EANN design process and the high
generalization capacity if compared with the trial and
error technique.
CONCLUSIONS
In this work was presented a computer tool which
analyzes the spectrometric and dosimetric information
obtained by means of the EANN technology by using the
software NGO. The results obtained in this work, reveal
that the hybrid technology of EANN in the neutron
spectrometry and dosimetry domain of the nuclear
sciences it is a very convenient approach for the ANN
design if compared with the trial and error technique used
for the researchers.
NSD-TB has the capability to analyze spectrometric
and dosimetric information obtained with several neutron
unfolding techniques as for example: iterative procedures
and the approaches based on ANN or EANN, as is the
case of the present work.
According to the results obtained in this work, more
research is needed in order to improve the EANN
performance.

3.
4.

H.R. Vega-Carrillo, H.-D. V.M., M.-A. E., M.-S.

G.A., I. de la Torre M.P., B. R., P.-F. S., M.-V. R.,
A.-A. T., and O.-R. J.M. Neutron spectrometry using
artificial neural networks. Radiation Measurements,
41:425431 (2006).
J.M Ortiz-Rodriguez, M.R. Martinez-Blanco and
H.R. Vega-Carrillo. Robust design of artificial
neural networks applying the Taguchi methodology
and DoE. Procc. Electronics, Robotics and
Automotive Mechanics Conference (CERMA),
Cuernavaca, Morelos, Mexico, 26-29 September,
Vol. II, 131-136, IEEE Computer society, (2006).
X. Yao. Evolutionary artificial neural networks.
International Journal of Neural Systems, 4(3):203
222 (1993).
J.M Ortiz-Rodriguez, M.R. Martinez-Blanco and
H.R. Vega-Carrillo. Artificial neural networks
modeling evolved genetically, a new approach
applied in neutron spectrometry and dosimetry
research areas. Procc. Electronics, Robotics and
Automotive Mechanics Conference (CERMA),
Cuernavaca, Morelos, Mexico, 30 September 3
October,Vol. 0, 87-392, IEEE Computer society,
(2008).

2009 Electronics, Robotics and Automotive Mechanics Conference

NSDann, a LabVIEW tool for neutron spectrometry and dosimetry based on the
RDANN methodology
Martinez-Blanco M.R.2 , Ortiz-Rodriguez J.M.1,3 and Vega-Carrillo H.R1,2,3
Universidad Autonoma de Zacatecas
U. Academicas: 1-Ingeneieria Electrica y 2-Estudios Nucleares
Avda. R.L. Velarde No. 801, Col. Centro, C.P. 98000, Zacatecas, Mexico
3-Universidad de Cordoba en Espana
Depto. de Electrotecnia y Electronica, Escuela Politecnica Superior
Avda. Menendez Pidal, s/n - Escuela Politecnica Superior, Cordoba, Espana
mrosariomb@yahoo.com.mx
Abstract

types of functions that characterize to the human thought.

The search of the AI has taken two fundamental roads: the
psychological and physiologic researching of the human
thought nature, and consequently the technological development of more and more complex computer systems. In
this sense, the term AI has been applied to systems and computer programs able to carry out complex tasks, simulating
the operation of the human thought, although still very far
from this [20].
The aim of the AI is the development of paradigms or
algorithms that require machines to develop cognitive tasks,
in those that the humans are usually better. A system of AI
should be able to make three things: to store knowledge, to
apply the knowledge stored in the solution of problems and
to acquire new knowledge by means of the experience, this
is, a system of AI has three key components: representation,
reasoning and learning [20].
At the present, diverse tendencies exist for the development of AI systems such as: expert systems (ES), genetic
algorithms (GA) [8, 9], fuzzy logic (FL) [18], artificial neural networks (ANN) [14, 19], etc. The ES try to reproduce
the human reasoning in a symbolic way. The GA use the
Darwininan theory of the survival and ANN make it from
a more biological perspective recreating the structure of a
human brain.
The theory of ANN has offered an alternative to the
classic calculation for those problems in which the traditional methods have not given very convincing or not very
convenient results. The most successful applications in
ANN are: image and voice processing, recognition of patterns, planning, adaptive interfaces for man-machine systems (robotics), prediction, optimization and control and
recently has been applied in the neutron spectrometry and
dosimetry research area of the nuclear sciences, by means

Nowadays the ANN theory it is offering an alternative for

those problems in which the traditional methods have not
given very convincing or not very convenient results. The
most successful applications in ANN are: planning, prediction, optimization and recently has been applied in the neutron spectrometry and dosimetry research area. The use of
the ANN technology is a useful alternative to solve the neutron spectrometry and dosimetry problems, however, several
drawbacks must be solved in order to simplify the use of
these procedures. In the present work was developed under
the LabVIEW programming enviroment, a computer tool
based on AI technology. Starting from the extracted knowledge of an ANN designed by means of the Robust Design of
Artificial Neural Networks methodology, which is capable
to unfold neutron spectra and to simultaneously calculate
13 equivalent doses. The designed tool presents the following characteristics: was designed in a users graphic interface, it is easy to use, is not essential on the part of the user,
to have knowledge of ANN neither of neutron spectrometry
and dosimetry to operate the tool, the execution speed and
performance of the program is high and, contrary to classical existent codes based on iterative algorithms, in the application developed in this work, is not required an initial
guess spectrum to obtain the neutron spectra wanted. Finally, this tool needs as entrance data just the count rates
coming from the BBS.

1. Introduction
In their wider sense, the term Artificial Intelligence (AI),
indicates the capacity that has a device to carry out the same
978-0-7695-3799-3/09 $26.00 2009 IEEE
DOI 10.1109/CERMA.2009.68

102
131

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:05 from IEEE Xplore. Restrictions apply.

of the Bonner spectrometric system (BSS) [4, 11, 16, 15].

However, several drawbacks are present with the use of the
ANN technology and more research is needed in order to
overcome them.

1.1

through the reconstruction of the spectrum. During the last

years have been carried out intents to develop new codes of
neutron spectra unfolding like: MITOM [23], FRUIT [3],
BUMS [22], etc., all of them based on iterative unfolding
algorithms. However, still present the serious drawback of
requiring a very expert user for their operation and the necessity to provide an initial guess spectrum for the deconvolution of the spectrum.
These considerations together with those suggested by
the experience in the radiological protection around the facilities, suggests that a reconstruction code dedicated to the
neutrons monitory in operational workspaces would be of
great help for the community that is devoted to the radiological protection [3]. To solve the mentioned problems
the researchers are studying and proposing alternative methods, which have as purpose to carry out an efficient neutron
dosimetry. Inside these alternative techniques, several unfolding procedures have been reported, combined with several types of experimental methods, based in AI technology,
such as: GA, ANN and Evolutive ANN [6].
Recently, the use of the ANN technology has been applied with success in the neutron spectrometry and dosimetry problems, however, the learning and structural parameters determination of the ANN still reamains a serious problem for the researchers. The learning and structural parameters are factors that contribute in a significant way with the
performance of the net that is beeing designed, and it has
been observed that the researchers in the nuclear sciences
area, carry out the selection of this parameters through the
trail and error technique, which produces ANN of poor performance and low generalization capacity.
For the ANN design, the net topology should be determined, however, many factors affect the ANN training,
learning and performance. Over the past years new proposals of ANN design have been made because the serious
drawbacks due to the trial and error thecnique. Recently, a
novel approach known as Robust Design of Artificial Neural Networks (RDANN) Methodology has been applied in
the design process of ANN [5]. The RDANN methodology
is a systematic and methodological aproach of ANN design
based on statistical methods which maximize the ANN performance and their capacity of generalization.
The RDANN methodology was used whith success in
the ANN parameter determination of the neutron spectrometry and dosimetry problems, which has become in an alternative procedure in the research area of nuclear sciences.
However, it has been observed that due to the novel of such
proposals there are no tools for the end user, which allow
to carry out in an efficient way the neutron spectra unfolding and dose calculus, and with it to simplify the procedures
and answer times for obtaining the spectrum and the wanted
dose.
The aim of the present work was to develop a com-

Neutron spectrometry by means of the BSS

Among all the available techniques for the neutron spectrometry, the BSS has been the more used for the purposes
of the radiological protection [23], due to the advantageous
characteristic that this system presents [2]. With the BSS,
which is a thermal neutron detector that is located at the
center of a high-density polyethylene spheres whose diameters are 2, 3, 5, 8, 10, 12, 16 and 18 inches [10], neutron
spectrum from thermal to several MeV can be obtained. The
BSS response matrix, the count rates and the neutron spectrum are related through the Fredholm integro-diferential
equation of the first type, whose discrete version is.
Cj =

n
X

Ri,j i , i = 1, 2, ..., m

(1)

i=1

where Cj is the j-th counts rate, Ri,j is the j-th detector

response to neutrons at the i-th energy interval, i is the
neutron fluence whithin the i-th energy interval and m is
the number of spheres utilized.
The most delicate part in the neutron spectrometry based
on the BSS, is the unfolding process. The unfolding spectra
of the neutrons measured consist on establishing the rate of
energy distribution of fluency (E), known as the response
matrix, Ri,j , and the group of carried out measures, Cj .
Because the number of incognits overcormes to the number
of equations, this is an ill-conditioned system and has an
infinite number of solutions. The procedure of selecting the
solution that has meaning for the problem type, is part of
the unfolding process. [17, 25, 24].
When using the integral counts of the BSS it is possible
to unfold the neutron spectrum, using diverse methods such
as: Monte Carlo, parametrization and iterative procedures.
However, the weight, the unfolding spectra procedure, the
necessity of a very expert user, the high time consumming,
the necessity to use an unfolding code (BUNKI, SAND,
etc.) and the resolution of the spectrum, are some of the
problems that this system presents. Each of the mentioned
difficulties have motivated the development of alternative
techniques, some of them based on AI technology such as
GA [21]or ANN [16, 4].

1.2

Neutron spectrometry by means of ANN

The current interest in the neutron spectrometry problem, has stimulated the development of diverse unfolding
procedures which try to obtain a better energy resolution

103
132

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:05 from IEEE Xplore. Restrictions apply.

puter tool for the end user, under the LabVIEW programming environment, starting from the extracted knowledge
of an ANN designed by means of the RDANN methodology, which is capable to unfold neutron spectra and to simultaneously calculate 13 equivalent doses. The designed
tool presents the following characteristics: was designed in
a users graphic interface, it is easy to use, is not essential
on the part of the user, to have knowledge of ANN neither
of neutron spectrometry and dosimetry to operate the tool,
the execution speed and performance of the program is high
and, contrary to the classical existent codes based on iterative algorithms, in the application developed in this work, is
not required an initial guess spectrum to obtain the wanted
neutron spectra. Finally, this tool needs as entrance data just
the count rates coming from the BBS.

information and build the main program (front end of the

final user) of the application.

Figure 1. Knowledge extraction of the ANN

In order to design the computational tool based on ANN
technology, the neural network structure is defined as a collection of parallel processors connected togheter in the form
of a direct graph, organized such that the network structure
solves the problem being considered. Referring to figure
2, as a typical network diagram, each processing element
in the network can be represented schematically as a node,
whith connections between units.

2. Materials and Methods

As was mentioned before, in previous works a new approach known as RDANN methodology was used to build
several net topologies able to solve the neutron spectrometry and dosimetry problems [6]. The parameters of the best
net designed are: 7 neurons in the entrance layer, corresponding to the count rates taken from the BSS, 26 neurons
in the first hidden layer, 39 neurons in the second hidden
layer and 44 neurons in the output layer, corresponding 31
of them to the neutron spectra unfolded and the remaining ones to the equivalent doses [7]. Unfortunatelly, the
end user do not have access to a computer system designed
specifically to obtain the neutron spectra and the equivalent
doses. For the anterior, was extracted the knowledge stored
at synaptic weigths of the mentioned ANN, and a graphical
tool was builded for the end user, based on this information.

2.1

ANN Structure and knowledge extraction

Figure 2. Multilayer ANN schema

Most of ANN models share the basic concepts of distributed and highly interconected processing elements (neurons). The computer tool proposed in this work, will be
builded on these simple concepts, implementing an interconnection scheme and structure similar to the ANN designed with the RDANN methodology in order to design
an application oriented to the end user, showed in figuren
2, capable to solve the neutron spectrometry and dosimetry
problems in a frendly graphical user interface, easy to use
and with high speed of operation and performance.
As shows figure 1, the general design process consists on
extract the information stored at sinaptic weights matrices
of the ANN designed and then build a framework under the
LabVIEW programming environment, whith the structure
of the ANN previously designed. Once the information was
extracted and the framework was builded, the next step will
be to introduce on the ANN framework the sinaptic weight

In this ANN scheme, the processing elements (neurons)

are grouped into layers. A layer is a collection of neurons. The layer receives differen names, according with its
location in the network. The input layer, receives the input signals of the network (p1 , p2 , ..., pR ). The hidden layers, dont have contact with the exterior, their elements can
have different connections which determine the net topology, however, a serious drawback is to determine the proper
selection of the optimum number of hidden layers and nodes
per layer for each particular problem, which usually is made
through the trail and error technique. The output layer, receives the information from the hidden layers and transmit
the response to the exterior (a1 , a2 , ..., aS ). If a net with
several layers is considered,
P each layer will have their own
weight matrix W , adders , their own bias vector b, an in

put vector
p , a transfer function f and an output vector
a
133
104

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:05 from IEEE Xplore. Restrictions apply.

[1].

2.2

dot product of the weight vectors and input vector. The dot
product Wij Yj (for j = 1,2...n) is the net input to the i-th
neuron. Also the output of the neuron could be an input to
some other neuron through another set of weights.
In Figure 4, the dot product of the weight vector and the
input vector is performed. This is the net input to the neuron in which a sigmoid activation function is applied to the
net input of the neuron. The Sigmoid function is a subVI, which is similar to function-call or sub-routine in text
based programming languages. The main library function
that is used here is the Dotproduct.VI, which performs the
dot product of two vectors.

LabVIEW Implementation of NSDann

Applications created with LabVIEW, a powerful graphical programming software developed by National Instruments, are refered as Virtual Instruments (VI) [12, 13]. VI
source code is created using the graphical programming
languaje G in a window called the block diagram. Input
and output interfacing with the VI is performed in another
window called the front panel. The graphical icon based
source code and interfacing creates very user-frendly applicationsand and eliminates typing in lenghtly characterbased code.
LabVIEW has the unique ability to develop data flow
diagrams that are highly parallel in structure. So LabVIEW
seems to be a very effective approach for building neural
nets based applications, and has been used here for building
a computer tool for neutron spectrometry and simultaneous
dosimetry based on ANN technology designed by means of
the RDANN methodology.
In order to build the computer tool denominated Neutron
Spectrometry and Dosimetry By means of Artificial Neural
Networks (NSDann) whithin LabVIEW programming environment, the neuron, the basic building block of an ANN
was examined. As shows figure 3, neurons are processing
elements that accept signal inputs Yj , process them using a
function f and then form an output signal Yi . Each input to
a neuron has a weight or gain associated with it.

Figure 4. Simple neuron model in LabVIEW

The next stage in the design process of the ANN structure was to design a model of a layer of neurons. Whith the
general model of the neuron, was designed the LabVIEW
block diagram showed in Figure 5, which is a general programming structure of a layer of neurons.

Figure 3. Simple artificial neuron model

The neuron output Yi is a function of the weighted input
that can be expressed by:
n
X
f(
Wi,j Yj )

Figure 5. ANN layer design in LabVIEW

From figure 5, it is assumed that there are 25 neurons
in a hidden layer. The net inputs to the 25 neurons in the
hidden layer are obtained by using the A* Vector.VI which
performs the dot product of each weight vector in the weight
matrix and the input data vector. Following that, a for loop
is used to apply the non-linear sigmoid activation function
(Sigmoid.VI) to these net inputs. The enable indexing option available to the for loop could be used to apply the sigmoid function to each of those 25 net inputs. Even the output of the for loop could be indexed so that one obtains 25
hidden layer outputs in the form of a 25 x 1 vector.

(2)

j=1

where Wi,j are the weights associated with the i-th neuron, n is the number of neurons and Yj are the inputs to the
neuron. Yi is the output of the ith neuron.
Taking in consideration the examination made for a single neuron, as first step in the design process of the NSDann
computer tool, was designed the basic model of a single
neuron configuration under LabVIEW, showed in figure 4,
where can be seen that a sigmoid function is applied to the

134
105

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:05 from IEEE Xplore. Restrictions apply.

Once the layer of neurons was created, the next step was
to build the general model of the ANN designed by means
of the RDANN methodology, capable to solve the neutron
spectrometry and dosimetry problems. For the anterior, the
knowledge stored at synaptic weights of the ANN was extracted and inserted in the framework designed according
to the structure of the designed network topology, as shows
figure 6.

computer tool is that for the calculation of the equivalent

doses is not required the information of the neutron spectra,
contrary to the unfolding codes used at present which use
the neutron spectra and the dose conversion factors for calculating the equivalent doses. NSDann calculate the equivalent doses independiently of the neutron spectra and in a
simulteneous time because the original training of the network.

Figure 6. Block diagram of NSDann

From figure 6, can be seen that according the net topology described before, in the general framework which contains the synaptic weights extracted to the ANN previously
designed, there are four layers. The input layer are the seven
count rates taking from the BSS. The first and second hidden layers have 26 and 39 neurons respectively. Finaly, the
output layer has 44 neurons as was explained before. In this
figure can be observed that was programmed an aditional
routine which graph the neutron spectra and doses calculaed
by means of the ANN structure programmed on LabVIEW.
As next step, was designed the front end graphical user
interfase of the NSDann computer tool, showed in figure
7, where can be seen that the NSDann computer tool just
requires as entrance data seven count rates comming from
a BSS, showed on the left side of the main window. After writing this information and pressing the Spectra unfold
button, after a few seconds, the application graph the neutron spectra unfolded as well as the equivalent doses calculated, showed at the center of the main window. At the
same time, the numerical information calculated, is showed
on the right side of the main window.

Figure 7. 239 P uBe neutron spectra unfolded

and 13 doses calculated whith NSDann
In a similar way, in figure 8 is showed the neutron spectra unfolded and its corresponding 13 equivalent doses of a
241
AmBe neutron source.

3. Results
In figure 7, is showed the neutron spectra unfolded
and the simultaneous calculation of 13 equivalent doses by
means of the NSDann computer tool, for a 239 P uBe neutron source. In the right side of the main window, the end
user writes the counts rate readed from a BSS, then pressing
the Spectra unfold button, after a few seconds the neutron
spectra is unfolded and simultaneously 13 equivalent doses
are calculated. An interesting characteristic of the NSDann

Figure 8. 241 AmBe neutron spectra unfolded

and 13 doses calculated whith NSDann

106
135

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:05 from IEEE Xplore. Restrictions apply.

4. Conclusions

[7] CERMA. A New Computer Tool Based on ANN Technology, Applied in Neutron Spectrometry and Dosimetry Research Areas, volume 00. Proceedings of the 2008 Electronics, Robotics and Automotive Mechanics Conference, 2008.
[8] D. L. Chambers. Handbook of genetic algorithms. Applications, volume 1. CRC Press, 2001.
[9] D. E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison Wesley, 1989.
[10] P. Goldhagen. Bonner-sphere neutron spectrometry, 1997.
[11] V.-C. H.R., H.-D. V.M., M.-A. E., M.-S. G.A., I. de la
Torre M.P., B. R., P.-F. S., M.-V. R., A.-A. T., and O.-R.
J.M. Neutron spectrometry using artificial neural networks.
Radiation Measurements, 41:425431, 2006.
[12] N. Instruments. Manual de referencia de funciones y iv.
[13] N. Instruments. Manual de referencia de programacion en
g.
[14] M. J. Jain A. K. and M. K. M. Artificial neural networks: a
tutorial. IEEE: Computer, 29(3):3144, 1996.
[15] K.-F. R. S. S. Kardan M. R. and G.-N. M. Fast neutron
spectra determination by threshold activation detectors using neural networks. Radiation Measurements, 38:185191,
2004.
[16] K.-F. R. S. S. Kardan M. R. and G.-N. M. Neutron spectra unfolding in bonner spheres spectrometry using neural
networks. Radiation Protection Dosimetry, 104(1):2730,
2004.
[17] S. S. K.-F. R. Kardan M. R. and G.-N. M. Neutron spectra unfolding in bonner spheres spectrometry using neural
networks. Radiation Protection Dosimetry, 104(1):2730,
2003.
[18] C. J. Lakhmi. Fusion of neural networks, fuzzy sets, and genetic algorithms: industrial applications. 1st edition, 1998.
[19] R. Lippmann. An introduction to computing with neural
nets. IEEE ASSP Magazine, 4(2):422, 1987.
[20] M.-B. M.R. Espectrometra de neutrones y calculo simultaneo de dosis equivalentes empleando la metodologa
de diseno robusto de redes neuronales, 2006.
[21] B. Mukherjee. Andi-03: a genetic algorithm tool for the
analysis of activation detector data to unfold high-energy
neutron spectra. Radiation Protection Dosimetry, 110(14):249254, 2004.
[22] H. N. Sweezy J. and V. K. Bumsbonner sphere unfolding made simple: an html based multisphere neutron spectrometer unfolding package. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 476(1-2):263
269, 2002. doi: DOI: 10.1016/S0168-9002(01)01466-8.
[23] F. F. B. M. Tomas M. and M. H. Mitom: a new unfolding code based on a spectra model method applied to neutron spectrometry. Radiation Protection Dosimetry, 110(14):545548, 2004.
[24] H.-D. V. M. M.-A. E. G. E. L. A. Vega-Carrillo H. R. and
I. M. P. Artificial neural networks technology for neutron
spectrometry and dosimetry. Radiation Protection Dosimetry, 126(1-4):408412, 2007.
[25] H.-D. V. M. M.-A. E. M. S. G. A. A. A. T. Vega-Carrillo H.
R. and O. R. J. M. Neutron spectra unfolding with artificial neural networks. ENINVIE 2005, Encuentro de Investigacion en Ingeniera Electrica, pages 132139, 2005.

The neutron spectra unfolding and the determination of

the equivalent doses, are not well conditioned problems
whose solution is carried out by means of codes of free distribution, where the obtained results turn out not to be reliable because the same ones depend on the experience of the
experimenter and of the quality of an a priori available
answer (the initial guess spectrum).
The neutron spectra unfolding measured by means of the
BBS system and making use of the ANN technology is an
alternative technique, which is being used at the present
time, because the better results obtained in comparison to
the classical methodologies, however, it has been observed
that due to the novel of such proposals there are no tools for
the end user, which allow to carry out in an efficient way the
neutron spectra unfolding and dose calculus, and with it to
simplify the procedures and answer times for obtaining the
spectrum and the wanted dose.
In this work a new computer tool based on AI technology was designed by means of the knowledge extracted
to an ANN designed with the RDANN methodology in
prior works in order to solve the neuton spectrometry and
dosimetry problems in a graphical user interface oriented
to the end user. The computational code which takes as entrance data just the counts rate comming from a BSS, constitutes a compact tool able to be used in the characterization
of neutron fields in nuclear facilities as for example: nuclear centrals, hospitals, research centers, neutron sources
of industrial use, etc.

References
[1] Artificial neural networks matlab tool box, ver.7.0, 2007.
[2] C. M. H. J. B. T. D. J. Alevra A. V. and S. H. Experimental
determination of the response of four bonner sphere sets to
monoenergetic neutrons (ii). Radiation Protection Dosimetry, 40(2):91102, 1992.
[3] D. C. E. A. Bedogni R. and F. F. Fruit: An operational
tool for multisphere neutron spectrometry in workplaces.
Nuclear Instruments and Methods in Physics Research A,
580(3):13011309, 2007.
[4] D. M. S. Braga C. C. Application of neural networks for
unfilding neutron spectra measured by means of bonner
spheres. Nuclear Instruments and Methods in Physics Research Section A, 476(1-2):252255, 2002.
[5] CERMA. Robust design of artificial neural networks applying the Taguchi methodology and DoE. Proceedings of
the 2006 Electronics, Robotics and Automotive Mechanics
Conference, 2006.
[6] CERMA. Artificial Neural Networks Modeling Evolved Genetically, a New Approach Applied in Neutron Spectrometry
and Dosimetry Research Areas, volume 00. Proceedings of
the 2008 Electronics, Robotics and Automotive Mechanics
Conference, 2008.

107
136

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:05 from IEEE Xplore. Restrictions apply.

2009 Electronics, Robotics and Automotive Mechanics Conference

A Computational Tool Design for Evolutionary Artificial Neural Networks in

Neutron Spectrometry and Dosimetry
Ortiz-Rodriguez J.M.1,3 , Martinez-Blanco M.R.2 , Gallego Eduardo4 and Vega-Carrillo H.R1,2,3
Universidad Autonoma de Zacatecas
U. Academicas: 1-Ingeneieria Electrica y 2-Estudios Nucleares
Avda. R.L. Velarde No. 801, Col. Centro, C.P. 98000, Zacatecas, Mexico
3-Universidad de Cordoba en Espana
Depto. de Electrotecnia y Electronica, Escuela Politecnica Superior
Avda. Menendez Pidal, s/n - Escuela Politecnica Superior, Cordoba, Espana
4-Universidad Politecnica de Madrid
Depto. de Ingenieria Nuclear
Jose Gutierrez Abascal, 2, E-28006, Madrid, Espana
morvymm@yahoo.com.mx
Abstract

tational model of adaptive systems, i.e, Evolutionary Artificial Neural Networks (EANN) [33], to utilize the learning
power of neural networks and adaptive capabilities of evolutionary algorithms. EANN refer to a special class of ANN
in which evolution is another fundamental form of adaptation in addition to learning. EANN can be regarded as a
general framework for adaptive systems, i.e., systems that
can change their architectures and learning rules adaptively
whithout human intervention. The evolution and learning
of EANN make their adaptation to a dynamic environment
much more effective and efficient.

Evolutionary Artificial Neural Networks have been studied widely in recent years. They provide not only an automatic method to design ANN, but also an approach to study
evolution and learning in the same framework. In prior
works was analyzed the intersection of artificial neural networks and evolutionary algorithms applied in the neutron
spectrometry and dosimetry research area, however, several
drawbacks were encountered, due mainly to the novel of this
approach. Because this is an emergent technology, a lack
of tools exists to carry out the analysis of the obtained results. In this work, it is presented the design of a computer
tool which allows to analyze the neutron spectromtery and
dosimetry results obtained by means of the hybrid technology based on evolutionary artificial neural networks. This
tool was designed under the LabVIEW programming environment in a graphic user interface, which make easy, intuitive and friendly their use. At the same time,the operation
speed of the program is high, carrying out the analysis in
seconds.

1.1

Artificial neural networks (ANN) have been used widely

in many application areas in recent years. A network consists of a number of network nodes connected by links,
and the specific pattern of connections defines the networks topology [16, 15, 19]. Most applications use feedforward ANN and the backpropagation (BP) training algorithm. There are numerous variants of the classical BP training algorithm and other training algoritms which assume a
fixed ANN architecture. They only train weights in the fixed
architecture that includes both connectivity and node transfer functions. Although the BP learning algorithm, provides
a method to train feedforward ANN, it is not free of problems. Many factors affect the learning and they should be
treated to have a process of successful learning. The problem of designing a near optimal ANN architecture for an
application remains unsolved, however this is an important
issue [33].

1. Introduction
Evolution and learning are two fundamental forms of
adaptation [32, 22]. Artificial neural networks (ANN) [12,
34, 23, 10] and evolutionary algorithms (EA) [25, 29, 13, 6]
are both abstraction of natural processes. Since early 1990,
they have been combined into a general and unified compu-

978-0-7695-3799-3/09 $26.00 2009 IEEE

DOI 10.1109/CERMA.2009.67

Evolutionary Artificial Neural Networks

113
84

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:03 from IEEE Xplore. Restrictions apply.

The design of a near optimal ANN architecture can be

formulated as a search problem in the architecture space
where each point represents an architecture [31]. Given
some performance criteria, e.g., minimum error, fastest
learning, lowest complexity, etc., about architectures, the
performance level of all architectures forms a surface in the
space. The optimal architecture design is equivalent to finding the highest point in the surface. According to the work
developed for Yao in 1993 [33] there are several characteristics whith such surface which make EA better candidates
for searching the surface.
Evolution can be introduced at various levels of ANN.
EA has been introduced into ANN at roughly three different levels: connection weights, architectures and learning
rules. The evolution of connection weights introduces an
adaptive and global approach to training, especially when
gradient information is unavailable or too costly to obtain
for gradient-based training algorithms to work. Architectural evolution enables ANN to adapt their topologies to
different tasks whithout human intervention. The evolution
of learning rules can be regarded as a process of learning
to learn in ANN where the adaptation of learning rules is
achieved trhough evolution.
There has been a strong interest in recent years in exploring these two forms of adaptation and their roles and
interactions in adaptive systems, in the solution of problems
of the real world. In the reasearch area of nuclear sciences,
the ANN technology has been used with success as an alternative approach in the solution of the neutron spectrometry
and dosimetry problems, however, still present some serious
drawbacks in the optimal determination of the net topology,
and researchers have mentioned the necesity to have new
alternatives in order to solve these problems. In prior works
was analized the evolution of ANN architectures where evolutionary algorithms were used to evolve ANN in the solution of the neutron spectrometry and dosimetry problems
[5].

1.2

cause neutron dosimetry is a complex technique [1] and

highly neutron energy dependent, and a precise knowledge on neutron spectrometry is highly essential for all
dosimetry-related studies as well as many nuclear physics
experiments. In consequence, it becomes necessary to develop additional measuring techniques to enhance the actual
workers monitoring systems.
The current interest in neutron spectrometry and dosimetry has stimulated the development of several deconvolution
procedures, based on Monte Carlo [17], maximum entropy
[30] and iterative algorithms [26, 1]. The neutron spectra unfolding and the dose equivalent calculation problems
by means of AI technologies such as GA [27, 28], ANN
[2, 18, 20, 11] and EANN [5], starting from the count rates
of the Bonner spectrometric system (BSS) [14, 21], is an alternative procedure on the nuclear sciences domain, which
has received great attention during the last years due to the
successful results that have been obtained.
ANN have received the higger attention among researchers, however, the use of this procedure is not free of
problems. Because this is an emergent technology, as the
investigations advance, new questions arise and new scientific and technological problems are presented. Many of the
previous studies in the neutron spectrometry and dosimetry
by means of ANN have found serious problems in the design process of the nets, i.e., the appropriate determination
of the structural and learning parameters of the neural networks [4]. EA and ANN are two relatively young research
areas that are subject to a steadily growing interest. As was
mentioned before, the structure of an ANN is a significant
contributing factor to its performance and the structure has
been generally heuristically chosen in most works on neutron spectrometry and dosimetry by means of ANN. The use
of EA as search techniques has allowed different properties
of ANN to be evolved in this research area.
Due to the above-mentioned, it is evident the necessity
that faces the scientific community of the nuclear research
area of having methods that allow to improve the applications developed by means of AI technologies and at the
same time the necesity to develop new technological tools
in order to solve the inconvenieces encountered along the
research carried out. In previous works was examined the
use of EA to optimize the structural and learning parameters
of ANN applied in the neutron spectrometry and dosimetry
problems, however, and because this is an emergent technology, was observed that there are no tools to carry out the
analysis of the obtained results, i.e. to prove and to validate
the neuttron spectra and doses obtained by means of EANN.
In this work, it is presented the design of a computational
tool designed under the LabVIEW programiming environment, that allows to analyze the spectra of neutrons and
equivalent doses obtained by means of EANN. This tool
was developed in a graphical user interface, is friendly, in-

Neutron spectrometry by means of EANN

In recent years, the characterization of ionizing radiation

fields in workplaces is one of the challenging activities in
radiation protection over the world [3, 24]. Nowadays, neutrons are widely used in many fields of both research and
technology [9]. The workers subject to these radiations especially those who are submitted to neutron risk have to be
well monitored and protected according to relevant national
regulations which are more and more restrictive. As a result, there is an increasing demand in the field of radiation
protection to quantify these various neutron fields and to
determine the radiation doses involved [7].
The dosimetry of neutron radiation is one of the most
complicated tasks in radiation protection [8] mainly be-

114
85

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:03 from IEEE Xplore. Restrictions apply.

tuitive and easy of operating, being high the performance

and speed of operation of the program.

ric information obtained by means of NGO or any other application for neutron spectrometry. After a few seconds, the
graphical and numerical comparison of the neutrons spectra
and doses calculated as well as the Chi square statistical test
are obtained.

2. Materials and Methods

In previous works, 300 EANN were designed, trained
and tested in a time of 00:01:43, selecting the best of all
of them [5]. In the EANN design process, was used the
software NeuroGenetic Optimizer (NGO). However, several drawbacks were observed with the use of NGO applied
in the solution of the neutron spectra unfolding and equivalent dose calculus problems. The first one was that NGO
has not the capability to realize statistical tests to validate
the obtained results. The second one was that NGO has
not the ability to separate the information of the unfolded
neutron spectra and the 13 equivalent doses calculated, to
be presented to the end user in a proper way. Because the
novelty of this methodology and the lack of tools for the
same, a computer tool was designed in order to overcome
the drawbacks before mentioned.
Figure 1, shows the main window of the program denominated Neutron Spectrometry and Dosimetry Toolbox
(NSDTB), where can be seen that the main window it is divided in two sections: The upper section shows the graphical (left side) and numerical (right side) information of the
neutron spectra unfolded. In a similar way, the lower section shows information of the equivalent doses calculated
by the EANN. As addtional element NSDTB has the ability to realize the Chi square statitical test among the spectra
and doses being analyzed.

Figure 2. Entrance file tex data

Figure 2, shows the opperation of the NSDTB computer
tool. After pressing the Load data button, a window appear,
requesting the entrance text file with the spectrometric and
dosimetric information wich will be analyzed.
Figure 3, shows the entrance text file tabulation separated. In order to edit this file, each column contains
the neutron spectrometric and dosimetric infrormation, obtained by means of NGO or any other spectrometric computer tool, corresponding the 31 fisrt rows to the neutron
spectra, and the remaining 13 to the equivalent doses. The
NSDTB is capable to analyze and compare simultaneuosly,
four neutron spectra and its corresponding equivalent doses
per file text.

Figure 1. Main window of NSDTB

Because NSDTB was designed in a graphical user interface under the LabVIEW programming environment, the
opertion of the tool is easy, intuitive and friendly to the end
user. Once the tool is initiated, the end user just require
pressing the Load data button, and a window will be open
as shows figure 2, requesting a text file, tabulated separated,
showed in figure 3, which contains the spectral and dosimet-

Figure 3. Entrance text file of NSDTB

115
86

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:03 from IEEE Xplore. Restrictions apply.

3. Results

desired data. From this figure, can be seen that after applying the statistical test, there are not statistical differences
between the desired and the obtained data.
Figure 7, shows the worst 13 equivalent doses and its
corresponding neutron spectra. In a similar way as in the
case 3, after applying the statistical tests, was observed that
there are not statistical differences among the desired and
obtained data.
From figures 6 and 7 can be seen that although the reported cases are the worst, the generalization capacity of
the EANN designed by means of NGO is high. However,
considering that the training time of the EANN was 0.34
seconds average, more training time is needed in order to
improve the performance and generalization capacity of the
best EANN selected.
The experiments described in the present work, demonstrate that the EA technology can be applied for finding the
optimal architectural and learning parameters of EANN in
order to solve the neutron spectrometry and dosimetry problems of nuclear sciences, reducing significantly the time
spent in the ANN design process itself if compared with
the trial and error approach, reported in the literature.

The NSDTB computer tool has the capability to compare in a graphical and numerical way, and at the same time
to analyze statistically by means of the Chi square test, the
neutron spectrometry and dosimetry information of several
unfolding procedures, just reading a text file which contains
such information. In this work, is analyzed the spectrometric and dosimetric information obtained with NGO by
means of EANN, as is showed in figures 4 through 7, where
can be seen a comparison of the the neutron spectra expected (blue line) against the obtained with NGO (red line)
and its corresponding 13 equivalent doses, after the training
and testing stages made in a time of 0.34 seconds average,
of the best EANN selected.
From figure 4 can be seen that there are graphical diferences among the spectral and dosimetric information obtained with NGO and the desired information, taken from
the testing data set. In this case, the neutron spectra unfolded is better than the equivalent doses calculated. However according the Chi square statistical test, can be seen
that there is not statistical difference between the neutron
spectra unfolded and the 13 equivalent doses desired and
obtained, because the value obtained for the Chi square test,
equal to 0.065 for the neutron spectra and 0.027 for the
equivalent doses, it is inferior to the critical value of the
Chi square tests, equal to 18.49 for the spectra and 5.23 for
the doses.

Figure 5. Best equivalent doses with NGO

Figure 4. Best neutron spectra with NGO

In a similar way for figure 5, is showed the better equivalent doses calculated and its corresponding neutron spectra obtained with NGO after the training and testing stages
of the better EANN. In this figure can be observed that
the equivalent doses are better than the neutron spectra unfolded, however the statistical tests show that there are not
differences between the spectra and doses desired and obtained.
Figure 6, shows the worst neutron spectra and its corresponding 13 equivalent doses calculated compared with the

Figure 6. Worst neutron spectra with NGO

87
116

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:03 from IEEE Xplore. Restrictions apply.

ciences. However, new problems arise of combining EA

and ANN, such as the proper determination of the EA parameters, and more research is needed in order to overcome
these drawbacks.

References
[1] D. C. E. A. Bedogni R. and F. F. Fruit: An operational
tool for multisphere neutron spectrometry in workplaces.
Nuclear Instruments and Methods in Physics Research A,
580(3):13011309, 2007.
[2] D. M. S. Braga C. C. Application of neural networks for
unfilding neutron spectra measured by means of bonner
spheres. Nuclear Instruments and Methods in Physics Research Section A, 476(1-2):252255, 2002.
[3] K. H. Brooks F. D. Neutron spectrometry fhistorical review and present status. Nuclear Instruments and Methods
in Physics Research A, 476:111, 2002.
[4] CERMA. Robust design of artificial neural networks applying the Taguchi methodology and DoE. Proceedings of
the 2006 Electronics, Robotics and Automotive Mechanics
Conference, 2006.
[5] CERMA. Artificial Neural Networks Modeling Evolved Genetically, a New Approach Applied in Neutron Spectrometry
and Dosimetry Research Areas, volume 00. Proceedings of
the 2008 Electronics, Robotics and Automotive Mechanics
Conference, 2008.
[6] D. L. Chambers. Handbook of genetic algorithms. Applications, volume 1. CRC Press, 2001.
[7] et al Mazrou H. Characterization of tue crna bonner spectrometer based on 6lii scintillator exposed to an 241ambe.
Radiation Measurements, 43:10951099, 2008.
[8] et al Souto E. B. Development of a brazilian-neutron
dosimeter. Nuclear Instruments and Methods in Physics Research A, 226:31743177, 2008.
[9] et al Wielunski M. Intercomparison excercise whith mev
neutrons using various electronic personal dosimeters. Radiation Measurements, 43:10631067, 2008.
[10] L. V. Fausett. Fundamentals of neural networks. Prentice
Hall, 1a edition, 1993.
[11] S. R. H. K. S. M. C. E. B. J. P. Fehrenbacher G. and
W. W. Proposal of a new method for neutron dosimetry
based on spectral information obtained by application of artificial neural networks. Radiation Protection Dosimetry,
83(4):293301, 1999.
[12] S. D. M. Freeman J. A. Neural networks: algorithms, applications, and programming techniques. Addison Wesley
Longman Publishing Co., Inc., 1991.
[13] D. E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison Wesley, 1989.
[14] P. Goldhagen. Bonner-sphere neutron spectrometry, 1997.
[15] J. L. Gupta M. M. and H. N. Static and dynamic neural
networks: from fundamentals to advanced theory. 2003.
[16] S. Haykin. Neural Networks: A Comprehensive foundation.
Prentice-Hall, 2a edition, 1999.
[17] V.-C. H.R., M.-A. E., O.-R. J.M., and A.-A. T. Neutron
spectra re-ginning and dose calculation using monte carlo
methods. Revista Mexicana de Fsica, 53:17, 2007.

Figure 7. Worst equivalent doses with NGO

4. Conclusions
In real world applications such as the encontered in the
neutron spectrometry and dosimetry problems by means of
ANN, a seriows drawback is the ANN design itslef. EA
seems to be a proper alternative to solve this problem. EA
can be used to assist in the ANN design and training, as
a novel approach in the nuclear sciences area, however the
lack of information and tools for the analysis of the results
obtained whith these new technologies, makes difficult the
work in this research area.
In this research, was presented a new computer tool designed in a graphical interface under the LabVIEW graphical programming environment, which analyze the information obtained whith EANN designed by means of NGO and
oriented to solve the neutron spectrometry and simultaneous dosimetry problems, using only the count rates measured with a Bonner spheres spectrometer system as entrance data.
The results obtained in this work reveals that the hybrid
technology of EANN applied in the neutron spectrometry
and dosimetry problems it is a very conevient approach of
ANN design in comparison with the classical approaches
based on the trial and error technique. According the results obtained in this work, it is evident that more research
is needed, in order to improve the EANN performance, because the combination of EA and ANN seems to be a very
useful approach.
The genetic search capability in ANN design is much
more effective than random searching approaches, as the
genetic process of recombining features vastly improves the
speed of identifying highly fit networks. It also has a potential advantage over just using personal experience in building neural networks, as new and potentially better solutions
may be found through this process than might be found using the nearly unavoidable assumptions made by the user.
The new ideas and concepts of EANN bring new life
into Artificial Intelligence research applied in the nuclear

117
88

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:03 from IEEE Xplore. Restrictions apply.

[18] V.-C. H.R., H.-D. V.M., M.-A. E., M.-S. G.A., I. de la

Torre M.P., B. R., P.-F. S., M.-V. R., A.-A. T., and O.-R.
J.M. Neutron spectrometry using artificial neural networks.
Radiation Measurements, 41:425431, 2006.
[19] M. J. Jain A. K. and M. K. M. Artificial neural networks: a
tutorial. IEEE: Computer, 29(3):3144, 1996.
[20] K.-F. R. S. S. Kardan M. R. and G.-N. M. Neutron spectra unfolding in bonner spheres spectrometry using neural
networks. Radiation Protection Dosimetry, 104(1):2730,
2004.
[21] R. M. A. B. Lacoste V. and M. H. Bonner sphere neutron spectrometry at nuclear workplaces in the framework of
the evidos project. Radiation Protection Dosimetry, 125(14):304308, 2007.
[22] C. J. Lakhmi. Fusion of neural networks, fuzzy sets, and genetic algorithms: industrial applications. 1st edition, 1998.
[23] R. Lippmann. An introduction to computing with neural
nets. IEEE ASSP Magazine, 4(2):422, 1987.
[24] S. B. R. L. McDonald J. C. and A. W. G. Neutron spectrometry for radiation protection purposes. Nuclear Instruments and Methods in Physics Research A, 476(1-2):347
352, 2002.
[25] M. Melanie. An Introduction to Genetic Algorithms. MIT
Press, 1998.
[26] S. C. Miller. AFITBUNKI: A Modified Iterative Code to
Unfold Neutron Spectra from Bonner Sphere Detector Data.
PhD thesis, 1993.
[27] B. Mukherjee. A high-resolution neutron spectra unfolding method using the genetic algorithm technique. Nuclear
Instruments and Methods in Physics Research A, 479(12):247251, 2002.
[28] B. Mukherjee. Andi-03: a genetic algorithm tool for the
analysis of activation detector data to unfold high-energy
neutron spectra. Radiation Protection Dosimetry, 110(14):249254, 2004.
[29] W. G. Periaux J. Genetic algorithms in engineering and
computer science. John Wiley & Sons, 1995.
[30] Z. A. Reginatto M. Bayesian and maximum entropy
methods for fusion diagnostic measurements with compact
neutron spectrometers. Review of Scientific Instruments,
79(2):023505+, 2008.
[31] F. Rothlauf. Representations for Genetic and Evolutionary
Algorithms. Springer, 2006.
[32] S. P. Whiteson S. Evolutionary function approximation for
reinforcement learning. The Journal of Machine Learning
Research, 7:877917, 2006.
[33] X. Yao. Evolutionary artificial neural networks. International Journal of Neural Systems, 4(3):203222, 1993.
[34] J. Zupan. Introduction to artificial neural network (ann)
methods: what they are and how to use them. Acta Chimica
Slovenica, 41(3):327352, 1994.

118
89

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on December 22, 2009 at 14:03 from IEEE Xplore. Restrictions apply.

Electronics, Robotics and Automotive Mechanics Conference 2008

Artificial Neural Networks Modeling Evolved Genetically, a New Approach

Applied in Neutron Spectrometry and Dosimetry Research Areas
Ortiz-Rodrguez Jos Manuel1,3, Martnez-Blanco Ma. del Rosario2, Gallego Eduardo4 and
Vega-Carrillo Hctor Rene1,2,3
Unidades Acadmicas de la Universidad Autnoma de Zacatecas:
1- Ingeniera Elctrica, 2 - Estudios Nucleares
Av. Ramn Lpez Velarde No.801, Col. Centro, 98000, Zacatecas, Mxico.
3 - Departamento de electrotecnia y electrnica, Escuela Politcnica Superior, Avda. Menndez
Pidal, s/n, Universidad de Crdoba, Crdoba Spain
4 - Depto. de Ing. Nuclear de la Universidad Politcnica de Madrid
C. Jos Gutirrez Abascal No. 2, 28006 Madrid, Spain
E-mail authors: morvymm@yahoo.com.mx, mrosariomb@yahoo.com.mx and,
fermineutron@yahoo.com

Abstract
1. Introduction
Recently, the use of the artificial neural networks
technology has been applied with success in the
research area of nuclear sciences, mainly in the
neutron spectrometry and dosimetry domains,
however, the structure (net topology), as well as the
learning parameters of the neural networks, are
factors that contribute in a significant way in the
networks performance. It has been observed that the
researchers in the nuclear sciences area carry out the
selection of the network parameters through the trial
and error technique, which produces poor artificial
neural networks with low generalization capacity and
poor performance. It has been observed that the use of
the evolutionary algorithms, seen as search and
optimization approaches, it has allowed to be possible
to evolve and to optimize different properties of
artificial neural networks, such as the proper synaptic
weight initialization, the optimum selection of the
network architecture or the selection of the training
algorithms. The aim of the present work is focused in
analyzing the intersection of the artificial neural
networks and the evolutionary algorithms, analyzing
like it is that the evolutionary algorithms can be used
to help in the design processes and training of an
artificial neural network, , in such a way that the
neural network designed is able to unfold in an
efficient way neutron spectra and to calculate
equivalent doses, starting only from the count rates
obtained from a Bonner spheres spectrometric system.

978-0-7695-3320-9/08 $25.00 2008 IEEE

DOI 10.1109/CERMA.2008.107

In last years, the Artificial Neural Networks (ANN)

technology has been utilized in the neutron
spectrometry and dosimetry research areas [1-6]. By
using the integral counts of the Bonner Spheres
Spectrometer (BSS) system it is possible to unfold the
neutron spectra [7-14]. The BSS response matrix, the
count rates and the neutron spectrum are related
through the Fredholm integro-diferential equation
which is an ill-conditioned equations system with an
infinite number of solutions. The weight, time
consuming procedure, the need to use an unfolding
procedure and the low resolution spectrum are the BSS
system drawbacks. Each of the mentioned difficulties
has motivated the development of complementary
procedures such as maximum entropy, GA [15, 16]
and ANN [1-6].
ANNs have been well recognized for its
approximation capability provided the input-output
data are available. Nevertheless, the conventional
training algorithm utilized in the model development is
always encountering difficulties to converge at global
solution. Many of the previous studies in neutron
spectrometry and dosimetry by using the ANN
approach have found serious drawbacks in the ANN
design process itself, mainly in the proper
determination of the structural and learning parameters
of the ANN, which are significant contributing factors
to its performance, because the optimal selection of

387

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:26 from IEEE Xplore. Restrictions apply.

these parameters follows in practical use no rules, and

their value is at most arguable, mainly because they
are generally heuristically chosen by using the trial and
error technique, which produces poor artificial neural
networks with low generalization capacity and poor
performance. In consequence more research has been
suggested in order to overcome these serious
drawbacks [1, 2].
In general, the development of a good ANN model
depends on several factors [17-23]. The first factor is
related to the data being used, the second factor is the
network architecture or model structure, the third
factor is the model size and complexity and finally, the
quality of the ANN model is also strongly dependent
on network training. This stage is essentially an
identification of model parameters that fits the given
data; and is perhaps the most important factor among
all. Training a neural network involves updating the
weights in such a manner that the error between the
outputs of the neural networks and the actual response
being modeled is minimized. Until today, many
researchers still prefer use the gradient search method
Back Propagation (BP) in training ANN. However,
this technique is a local search method and when
applied to complex nonlinear optimization problems,
can sometimes result in inconsistent and unpredictable
performances. One of the main hindrances is due to the
fact that searching of optimal weights is strongly
dependent on initial weights and if they are located
near local minima, the algorithm would be trapped; if
the initial guess of the ANN is near local maxima, it
will climb the gradient and get stuck. Several different
attempts have been proposed by various researchers to
alleviate the training problems. These include
imposing constraints on the search space, restarting
training at many random points, adjusting training
parameters and restructuring the ANN architecture.
However, some approaches are problem-specific and
not well accepted and different researchers tend to
prefer different methodologies.
Among the limitations of ANN, the followings
should be given added emphasis: network architecture,
training algorithm, training data, and process
relationship. All of these limitations have motivated
researchers to generate ideas of merging or hybridizing
ANN with other approaches in the search for better
performance. Some of the available schemes include
expert systems, statistical methods such as the Robust
Design of Artificial Neural Networks (RDANN)
methodology [24], fuzzy logic, wavelet transform and
as well as GA [25-28]. Aiming at improving the
accuracy and robustness of ANN model, GA [29-31]
was introduced in ANN modeling for connection
weights or architecture evolution. By evolving the

ANN some of the drawback of conventional ANN can

be overcome.
One of the more promising techniques to design the
structural and learning parameters of ANN is by
introducing adaptation of network training using GA.
The aim of the present work is focused in analyzing
the intersection of the artificial neural networks and the
GA, analyzing like it is that GA can be used to help in
the design processes and training of ANN, i.e., in the
optimum selection of the structural and learning
parameters of ANN, improving its generalization
capacity, in such a way that the neural network
designed is able to unfold in an efficient way neutron
spectra and to calculate equivalent doses, starting only
as entrance data from the count rates obtained of a
Bonner spheres spectrometric system.

2. Materials and Methods

Evolving Artificial Neural Networks refer to a
special class of ANN in which evolution is another
fundamental form of adaptation in addition to learning
[29-31]. In this case, GA can be used to adapt
connection weights, architecture design, learning rule
adaptation, input feature selection and so on. The
evolution of ANN can increase model adaptability to a
dynamic environment. In other words, the models
developed are more robust to dynamic nonlinear
process system. This benefits the ANN since its
generalization capability can be improved. In terms of
architecture design and learning algorithm, GA can
also help to reduce the uncertainties in some parameter
selection.
There are two major attempts to evolve ANN. The
first attempt is to conjugate GA with ANN by the
evolution of connection weights. The evolution
process can further divided into two stages. The first
stage only dealt with the offline set-up of the network
that GA was used in training of the connection weights
instead of using classical training algorithm. The
second stage extended the approach of off-line
training, by means of enabling on-line retraining
strategy to connections weights in a fixed topology
network. This increased the capability of ANN to
follow real-time interactive environment via
continuous updating the model knowledge from
retraining process data. The second attempt in
providing adaptability to the neural model is to evolve
its architecture. With the term architecture, intend not
only for the topological structure, but also the transfer
functions of the neurons. The evolution of
architectures allows neural network to adapt their
topologies to different tasks or to a non stationary

388

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:26 from IEEE Xplore. Restrictions apply.

changing environment without human intervention,

introducing an approach to automatic neural network
design
In this work, 300 genetically evolved neural
networks were designed, trained and tested in a time of
00:01:43, selecting the best of all of them. To train the
networks were used the count rates coming from 7
Bonner spheres as input and a data set of 128 neutron
spectra and 13 equivalent doses as outputs [1,2,32]. In
the analysis of the 128 data used in the training and
testing process of the ANN design, were used seven
neurons in the input layer (corresponding to the seven
Bonner Spheres used as input variables) and 44
neurons in the output layer (corresponding 31 of them
to the spectra to be unfolded and the 13 remaining to
the equivalent doses to be calculated). In the ANN
design process, was used the Root-Mean-Square Error
of Prediction (RMSEP) as objective function, which is
the square root of the sum of the squared differences
between the observed and predicted values for all
observations in the test set divided by the number of
such observations, to estimate the prediction error. To
design the optimal ANN was used the software
NeuroGenetic Optimizer (NGO), version 2.6, from
BioComp Systems, Inc. [33]. The original data were
automatically subdivided by the software into training
and test sets by using a random procedure. For the 128
neutron spectra and equivalent doses samples included
in the dataset, 112 samples were used as the
modelling/training set, and the remaining samples as
the test set.
According NGO, the processes of GA in the context of
ANN are the following. The GA process is based on a
fundamental cyclic process. This process consists of:
(1) Creating an initial population of genotypes
(genetic representations of the neural networks); (2)
Building neural networks (phenotypes) based on the
genotypes; (3) Training and testing the neural
networks to determine how fit they are; (4) Comparing
the fitness of the networks and keeping the best; (5)
Selecting those networks in the population which are
better, discarding those which arent; (6) Refilling the
population back to the defined size; (7) Pairing up the
genotypes of the neural networks; (8) Mating the
genotypes by exchanging genes (features) of the
networks; (9) Mutating the genotypes in some
random fashion; Then returning back to step (2) and
continuing this process until some stopping criteria is
reached or manually stops the process. Through the
process described previously, the better networks
survive and their features carry forward into future
generations and are combined with others to find better
and better networks for the problem considered. This
genetic search capability is much more effective than

random searching, as the genetic process of

recombining features vastly improves the speed of
identifying highly fit networks. The design parameters
of the ANN as well as the GA, used to evolve an ANN
able to unfold neutron spectra and to calculate
equivalent doses, by utilizing the NGO, were the
following: The GA parameters - generations run: 10,
population size: 30, selection was performed by the top
50% surviving, refilling of the population was done by
cloning the survivors, mating was performed by using
the TailSwap technique, mutations were performed
using the following technique(s): random exchange
technique at a rate of 25%. The ANN parameters used
were: the minimum network training passes for each
network were 20, the cutoff for network training
passes was 50, the input neural node influence factor
used was 0, the hidden neural node influence factor
used was 0, the limit on hidden neurons was 8, the
Root Mean Squared Error on training se was 1.4455,
the Min. Root Mean Squared Error on test set was
0.4002, the Root Mean Squared Error on validation set
was 0.00, the network type was a Fast-Back
Propagation neural network, the network employed
255 inputs and 1 hidden layer with 1 Logistic 1 Tanh 1
Linear neurons, there were 1 output neurons using the
Tanh transfer function.

3. Results
ANN have the potential to be used in the neutron
spectrometry and dosimetry research areas due to its
nonlinear mapping ability that enabled the capability to
accurately describe the behavior of complex systems.
The genetic search capability in ANN design is much
more effective than random searching approaches, as
the genetic process of recombining features vastly
improves the speed of identifying highly fit networks.
The general idea of combining GA and ANN was
the following: Information about the neural network
was encoded in the genome of the genetic algorithm.
At the beginning, a number of random individuals
were generated. The parameter strings were evaluated,
which means a neural network was designed according
to the genome information. Its performance was
determined after training with backpropagation.
Then, they were evaluated and ranked. The fitness
evaluation may take more into consideration than only
the performance of the individual. Finally, crossover
and mutation created new individuals that replaced the
worst - or all - members of the population. This
general procedure was quite straight-forward.
In this work, by using NGO for the genetically
evolved ANN modeling applied in neutron

389

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:26 from IEEE Xplore. Restrictions apply.

spectrometry and dosimetry, 34 training records at

training stage were selected at random by NGO, as is
showed in figure 1, where can be seen a comparison
between the desired neutron spectra (black line) and
predicted (blue line) of the optimal genetic ANN
designed. In figure 1, can be appreciated that the error
obtained in the optimal genetic ANN is small, which
implies that the ANN learning was good. Statistical
tests would be necessary to validate this inference.

Figure 2. A comparison between 9 testing records at testing

stage, of the optimal genetic ANN

Figure 1. A comparison between 34 training records at

training stage, of the optimal genetic ANN designed

In a similar way, in the testing stage, NGO selected

randomly 9 records to carry out another comparison
between the desired neutron spectra (black line) and
predicted (blue line) of the optimal genetic ANN
designed, with the purpose of checking the ANN
generalization capability, by using data that were not
included in the training stage, as shows figure 2, where
can be seen that as in the training stage case, the error
among the selected records is small, which is an
indicator that this ANN has good generalization
capability and in consequence this makes infer that this
ANN unfold neutron spectra of high performance.
Figure 3, shows a comparison among the curve of
errors obtained during the training and testing stages of
the best ANN evolved genetically. The vertical scale of
the figure shows the square root of the mean square
error, and the horizontal scale shows the number of
occasions in that the dataset was used to train the net.
As it is observed, at the beginning the error is big in
both cases, but it descends as the number of cycles is
increased with which the net was trained. In figure 3
can be appreciate that the error obtained in the training
stage (line in blue color), presents a soft curve if it is
compared with the error obtained in the test stage (line
in red color), which presents strong oscillations, which
implies that the testing stage was easier and better than
the training of modeling.

Figure 3. Root square mean error comparison at training and

testing stages of the optimal genetic ANN

The best genetic ANN was found on generation 9

after a runtime of 00:01:43. Once concluded the
training and testing processes of all the designed ANN
by means of NGO, it was observed that the best ANN
selected, it was even able to continue learning, the
anterior makes infer that if the AG-ANN parameters
used to evolve the net in question are modified, for
example by giving more training time and using more
times the data set training, the ANN performance as
well as the generalization capability can be
significantly improved, producing in consequence
better neutron spectra and calculated doses.

390

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:26 from IEEE Xplore. Restrictions apply.

A drawback was observed with the use of NGO

applied in the neutron spectrometry and dosimetry
domain, and is that the form of presenting the obtained
information to the nuclear researcher is not adequate,
since the information of the spectrum of neutrons and
the equivalent doses have to be separated and
presented in different graphics which not occurs in
NGO, because it presents the neutron spectra and dose
calculus in the same graphic, as is observed in figure 1
and 2. Because the tool explored in this work,
denominated Genetic Artificial Neural Networks
(GANN), although it is broadly well-known for the
investigators, but however, it is applied for first time in
the nuclear sciences area, specifically to solve both the
neutrons spectrometry and dosimetry problems, it is
observed that tools don't exist to solve the drawbacks
encountered with the use of NGO, reason why it is
necessary to carry out more investigation with the aim
of building tools that allow to solve the outlined
problems. At the present time are being carried out
spread works to propose solutions to the same ones
[34].

the system, a more practical approach is off-line set up

a rough model, followed by on-line update of the
model using GA. In this way, the merging of GA and
ANN will gain adaptability to dynamic environment
and lead to significantly better intelligent systems than
relying on ANN or GA alone. In other words, the GAANN model presented here should be more robust to
dynamic nonlinearity of the process involved.
Once concluded the training and testing processes
of all the designed RNA by means of NGO, it was
observed that the best RNA selected, it was even able
to continue learning, the anterior makes infer that if the
GA-ANN parameters used to evolve the net in
question, are modified, the ANN performance as well
as the generalization capability can be improved,
producing in consequence better neutron spectra and
calculated doses.
A serious drawback was observed with the use of
the NGO in the neutron spectrometry and dosimetry
solution problems, and is that it has not the capability
to separate the information of the unfolded spectra and
the 13 equivalent doses calculated, to be presented to
the end user in a proper way. Thus, the spectra and
dose calculus information is presented in a non
convenient way for its interpretation to the researcher.
Because the novelty of this methodology there are no
tools at this moment, to overcome the drawbacks
encountered in this research, being necessary to
develop scientific tools for such end. At present, work
is being carried out in this to alleviate this
inconvenience.
In this research, a new evolvable ANN modeling
approach called Genetic Artificial Neural networks
(GANN) methodology was presented for first time,
which utilizes an optimization process based on the
combination of genetic algorithms and artificial neural
networks applied in the design of an ANN, oriented to
solve the neutron spectrometry and simultaneous
dosimetry problems, using the count rates measured
with a Bonner spheres spectrometer system. The
experiments described in the present work,
demonstrate that the GA technology can be applied for
finding the optimal network parameters and
architecture for ANN, in the neutron spectrometry and
dosimetry domains of nuclear sciences.
In future works, would be necessary to perform
statistical tests to validate if exists statistical difference
between the desired data and the obtained.

4. Conclusions
In this work, GA technology by means of NGO was
used for first time, for the modeling and optimization
of an ANN capable to solve the neutron spectra
unfolding and simultaneous calculus of 13 equivalent
doses problems. The time spent in the ANN design
process and the performance obtained with the use of
this new methodology, is so inferior and at the same
time so superior, respectively, if compared with
heuristic approaches. [1-6]
The genetic search capability in ANN design is
much more effective than random searching
approaches, as the genetic process of recombining
features vastly improves the speed of identifying
highly fit networks. It also has a potential advantage
over just using personal experience in building neural
networks, as new and potentially better solutions may
be found through this process than might be found
using the nearly unavoidable assumptions made by the
user. The new ideas and concepts of GA and ANN
bring new life into Artificial Intelligence research.
However, the problem of combining GA and ANN lies
in the encoding of the network.
Unlike BP training algorithm of ANN, GA is a
global search algorithm based on the principle
survival of fittest. It simultaneously searches for
solutions in several regions, thus increasing the
probability of global convergence. Furthermore, since
it is impossible to formulate an a priori exact model of

5. References

391

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:26 from IEEE Xplore. Restrictions apply.

[16] Freeman, D.W., Edwards, D.R., Bolon, A.E., Genetic

algorithms-A new technique for solving the neutron
spectrum unfolding problem, Nucl. Instrum. Meth.
Phys. Res. A, 425 (1999).
[17] Fausett L., Fundamentals of neural networks, PrenticeHall, (1993).
[18] Gupta M.M., Lin J., and Homma N., Static and
Dynamic Neural Networks. From Fundamentals to
Advanced Theory, John Wiley & Sons, (2003).
[19] Haykin S., Neural Networks: A Comprehensive
Foundation, Prentice Hall, (1999).
[20] Jain, A.K., Mao J., and Mohuiddin, K.M., Artificial
neural networks: a tutorial, IEEE, 29 (1996).
[21] Lippmann R.P., An introduction to computing with
neural networks, IEEE ASSP MAGAZINE, 4 2 (1987).
[22] M.M. Gupta, J. Lin, and N. Homma, Static and
Dynamic Neural Networks from Fundamentals to
Advanced Theory, John Wiley & Sons, 2003.
[23] Kohonen, T., Procc. IEEE, 78 9 (1990).
[24] J.M. Ortiz Rodrguez, M.R. Martinez Blanco, and H.R.
Vega carrillo, Robust design of artificial neural
networks applying the taguchi methodology and DoE,
CERMA (2006).
[25] Goldberg, David E., Genetic Algorithms in Search,
Optimization, and Machine Learning, Addison-Wesley
(1989).
[26] Randy L. Haupt and Sue Ellen Haupt, Practical genetic
algorithms, Wiley (2004).
[27] Franz Rothlauf, Representations for Genetic and
Evolutionary Algorithms, Springer (2006).
[28] Mitchell Melanie, an Introduction to Genetic
Algorithms, MIT Press (1998).
[29] Xin Yao, a review of evolutionary artificial neural
networks, international journal of intelligent systems, 8
(1993).
[30] Lakhmi C. Jain; N.M. Martin, Fusion of Neural
Networks, Fuzzy Systems and Genetic Algorithms:
Industrial Applications, CRC Press (1998).
[31] S. Whiteson and P. Stone, Evolutionary Function
Approximation for Reinforcement Learning, journal of
machine learning research, 7 (2006).
[32] H.R.Vega-Carrillo, M.P. iguez, Catalogue to select
the initial guess spectrum during unfolding, Nucl.
Instrum. Meth. Phys. Res. A (2002)
[33] http://www.biocompsystems.com
[34] Martnez-Blanco Ma. del Rosario, Ortiz-Rodrguez Jos
Manuel and Vega-Carrillo Hctor Rene, A New
Computer Tool Based on ANN Technology, applied in
neutron spectrometry and dosimetry research areas,
accepted to be presented at CERMA (2008).

[1] H.R. Vega-Carrillo, V.M. Hernandez-Davila, E.

Manzanares-Acua, G.A. Mercado, E. Gallego, E.
Lorente, W.A. Perales, J.A. Robles, Artificial neural
networks in neutron dosimetry,Radiat. Prot. Dosim.,
118 251 (2006).
[2] H.R. Vega-Carrillo, V.M. Hernndez-Dvila, E.
Manzanares-Acua, G. Mercado, M.P. Iiguez, R.
Barquero, F. Palacios, R. Mndez, T. Arteaga, J.M.
Ortz, Neutron spectrometry using artificial neural
networks, Radiat. Measurem., 41 425 (2006).
[3] Braga, C.C. and Dias M.S., Application of neural
networks for unfolding neutron spectra neasured by
means of Bonner spheres, Nucl. Instrum. Meth. Phys.
Res., A 476 (2002).
[4] Cordes A., Fehrenbacher G., Schutz R., Sprunk M.,
Hahn K., Hofmann R., Biersack J.P., An approach to
unfold the response of al multi-element system using
artificial neural networks, Wahl W., IEEE Trans. Nucl.
Sci., 45 (1998).
[5] Fehrenbacher G., Schutz R., Hahn K., Sprunk M.,
Cordes, A., Biersack J.P., Wahl W., Proposal of a new
method for neutron dosimetry based on spectral
information obtained by application of artificial neural
networks, Radiat. Prot. Dosim., 83 (1999).
[6] Kardan M.R., Koohi-Fayegh R., Setaycshi S., GhiassiNejad M., Fast neutron spectra determination by
threshold activation detectorsusing neural networks,
Radiat. Measurem., 38 (2004).
[7] McDonald J.C., Siebert B.R.L., Alberts W.G., Neutron
spectrometry for radiation protection purposes, Nucl.
Instrum. Meth. Phys. Res., A 476 (2002).
[8] Thomas J. D., Neutron spectrometry for radiation
protection, Radiat. Prot. Dosim., 110 (2004).
[9] H.R. Vega-Carrillo, E. Manzanares Acua, J.M. OrtizRodrguez, and T. Arteaga Arteaga, Neutron spectra
re-binning and dose calculation using Monte Carlo
methods, Revista Mexicana de Fsica, 53 3 (2007).
[10] Matzke M., Unfolding procedures, Radiation
Protection Dosimetry, 107 (2003).
[11] El Messaoudi M, Chouak A, Lferde M, Cherkaoui R.,
Performance of three different unfolding procedures
connected to Bonner sphere data, Radiat Prot Dosim.,
108 (2004).
[12] Lacoste V., Gressier V., Pochat J.-L., Fernandez F.,
Bakali M., Boassoule T., Characterization of Bonner
spheres systems at monoergetic and thermal neutron
fields, Radiat. Prot. Dosim., 110 (2004).
[13] Vylet V., Response matrix of an extended Bonner
sphere system, Nucl. Instrum. Meth. Phys. Res., A 476
(2002).
[14] Vega-Carrillo,
H.R.,
Manzanares-Acua,
E.,
Hernndez-Dvila, V.M, and Mercado, G.A., Response
matriz of a multisphere neutron spectrometer with an
3He proportional counter, Rev. Mx. Fs., 51 (2005).
[15] Mukherjee B., A high-resolution neutron spectra
unfolding method using the genetic algorithm
technique, Nucl. Instrum. Meth. Phys. Res., A 476
(2002).

392

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:26 from IEEE Xplore. Restrictions apply.

Electronics, Robotics and Automotive Mechanics Conference 2008

A New Computer Tool Based on ANN Technology, applied in neutron

spectrometry and dosimetry research areas
Martnez-Blanco Ma. del Rosario2, Ortiz-Rodrguez Jos Manuel1,3 and Vega-Carrillo Hctor
Rene1,2,3
Universidad Autnoma de Zacatecas:
Unidades Acadmicas: 1- Ingeniera Elctrica, 2 - Estudios Nucleares
3 - Departamento de electrotecnia y electrnica, Escuela Politcnica Superior, Avda. Menndez
Pidal, s/n - Escuela Politcnica Superior, Crdoba Spain
E-mail authors: mrosariomb@yahoo.com.mx, morvymm@yahoo.com.mx and,
fermineutron@yahoo.com
The dosimetry of neutron radiation is one of the
most complicated tasks in radiation protection [1]. This
arises from several factors, related to either the
definition of the operational quantities, or the practical
problems in the instrument design and calibration [2].
Reliable determination of neutron doses in workplaces
is still an issue in the field of radiation protection [3].
Although there is a wide range of different devices
used for neutron spectrometry, the majority of the
instruments can be grouped together into a small
number of broad categories, each one based on a
common underlying technique [4-6]. Among the many
available neutron spectrometry techniques, the
multisphere or Bonner sphere spectrometer (BSS)
system is the most used for radiation protection
purposes [7-9], due to advantageous characteristics as
wide energy range (from thermal to GeV neutrons),
large variety of active or passive thermal sensors
allowing adapting the sensitivity to the specific
workplace, good photon discrimination and simple
signal management. Disadvantages are the poor energy
resolution, which does not allow appreciating fine
structures as narrow peaks, the weight, and the need to
sequentially irradiate the spheres, requiring, in general,
long exposure periods [1].
In the neutron spectrometry by means of the BSS
system and utilizing the classical approaches, as for
example Montecarlo, parametrization and iterative
procedures, the weight, time consuming procedure, the
need to use an unfolding procedure, the low resolution
spectrum, and the need to use the neutron fluence-todose conversion coefficients for calculating the
equivalent doses are some drawbacks which these

Abstract
In the neutron spectrometry and dosimetry research
areas by means of the Bonner spheres spectrometric
system utilizing classical approaches, such as
Montecarlo, parametrization and iterative procedures,
the weight, time consuming procedure, the need to use
an unfolding procedure, the low resolution spectrum,
and the need to use the neutron fluence-to-dose
conversion coefficients for calculating the equivalent
doses are some drawbacks which these approaches
offer. Each of the mentioned difficulties has motivated
the development of complementary procedures such as
maximum entropy, genetic algorithms and artificial
neural networks. The use of neural networks to unfold
neutron spectra and to calculate equivalent doses from
the count rates measured with BSS system has become
in an alternative procedure, which has been applied
with success, however, it is observed that doesn't exist
computer tools based on ANN technology to unfold
neutron spectra and to calculate equivalent doses. In
this work a customized front end user interface
software application, called Neutron Spectrometry and
Dosimetry by means of Artificial Neural Networks
(NSDANN) based on ANN technology, is presented,
which is capable to unfold neutron spectra and to
simultaneously calculate 13 equivalent doses, by using
only the count rates of a BBS system, in just a few
seconds.

1. Introduction

978-0-7695-3320-9/08 $25.00 2008 IEEE

DOI 10.1109/CERMA.2008.106

189

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:06 from IEEE Xplore. Restrictions apply.

approaches offer. Each of the mentioned difficulties

has motivated the development of complementary
procedures such as maximum entropy, Genetic
Algorithms (GA) and Artificial Neural Networks
(ANN). The use of ANN to unfold neutron spectra and
to calculate equivalent doses from the count rates
measured with BSS system has become in an
alternative procedure [10-15], which has been applied
with success, and some authors have expressed that
this technology merit attention because de success
obtained [2].
In the ANN design process, the choice of the
ANNs basic parameters (network topology, learning
rate, initial weights, etc) often determines the success
of the training process. The selection of these
parameters follows in practical use no rules, and their
value is at most arguable, because users have to choose
the architecture and determine many of the parameters
in a selected network by employing the trial and error
technique which produces ANN with poor
performance and low generalization capability
spending often large amount of time. To overcome
these drawbacks a new approach known as Robust
Design of Artificial Neural Networks (RDANN)
methodology has been proposed [22-27]. The RDANN
methodology is a powerful method based on parallel
processes where all the experiments are planned a
priori and the results are analyzed after all experiments
are completed. RDANN can help to provide answers to
critical design and construction issues of ANN.
As mentioned before, by using the integral counts
of the BSS system it is possible to unfold the neutron
spectra by using several methods like Montecarlo,
parametrization and iterative procedures, considered as
classical approaches. The use of ANN to unfold
neutron spectra and to calculate equivalent doses from
the count rates measured with BSS system has become
in an alternative procedure which has been applied
with success [10-15], however, the optimum ANN
parameter determination problem remains. RDANN
has been used to alleviate this drawback [22-27].
Provided a well-established response matrix and
adequate irradiation conditions, the most delicate part
of the BSS-based spectrometry are the unfolding
process. Many computerized BSS unfolding codes
have been developed, as for instance BONDI-97,
SAND, MAXED, BUNKIT [28]. Other than
conventional methods used to unfold the neutron
spectra, considered as new approaches, have recently
been developed such as MITOM [29], FRUIT [2],
GRAVEL [3,30] and NUBAY [3,30]. The critical
points of these neutron spectra unfolding codes are, in
general, the complexity of the codes, the need of a very
expert user and the need of realistic a priori

information, such as a default spectrum as close as

possible to the spectrum to be obtained [2]. These
considerations, together with those suggested by the
experience in radiation protection around neutron
producing facilities, suggest that an unfolding code
devoted to the operational workplace neutron
monitoring would be of great help to the radiation
protection community [2]. With this purpose, a new
computer tool based on ANN technology was
developed. In this work is presented a customized front
end user interface software application, called Neutron
Spectrometry and Dosimetry by means of Artificial
Neural Networks (NSDANN) based on ANN
technology, capable to unfold neutron spectra and to
simultaneously calculate 13 equivalent doses, by using
only as input data the count rates coming of a BBS
system, in just a few seconds if compared with the time
spent with the classical techniques, not being needed a
priori information about the spectra being calculated.

2. Materials and methods

In prior works, a large variety of neural network
architectures and topologies were developed, by using
the RDANN methodology, in order to build an ANN
capable to solve the neutron spectrometry and
dosimetry problems. The best ANN selected was:
network topology: 7:26.39:44, momentum: 0.1,
learning rate: 0.1, training algorithm: trainscg, target
mean square error: 1e-4, as is showed in figure 1,
where can be seen a basic scheme of the ANN
designed. In this figure, can be highlighted that the
ANN requires only as input data, the count rates
coming from a BBS system; the neutron spectra is
unfolded and 13 equivalent doses are simultaneously
calculated in an automated way, after just a few
seconds and in an independent way, i.e., for calculating
the equivalent doses, this ANN does not require the
spectrometric information.

Figure 1. Neutron spectrometry-dosimetry based on ANN

Once the networks parameters were determined

(which was the harder task), the same was trained and

190

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:06 from IEEE Xplore. Restrictions apply.

tested, however, from the literature reviewed, it was

observed that doesn't exists computer tools based on
ANN technology to make neutron dosimetry in nuclear
workplaces, in such a way that it exists the necessity
for building a computational tool for such an end,
which is the aim of this work.
The NSDANN computer tool was designed under
Matlab programming environment extracting the
numerical data stored at synaptic weights, and building
an end user interface as is showed in figure 2, where
can be seen that the NSDANN computer tool requires
as entrance data, only the measured Bonner spheres
counts rate which must be stored in an spreadsheet in a
local folder of the computer. After the counts rate are
reading by NSDANN, it produces the output, i.e., the
neutron spectra unfolded and 13 equivalent doses, after
a few seconds in a completely automated way. The
information produced by NDSANN is available in
both graphical and numerical format for a further
manipulation of the researcher.

response matrix of the UAB for 12 Bonner spheres and

a detector of the type 3He, whereas in figure 4 can be
seen the response matrix of the UAZ for 7 Bonner
spheres and a cm2 6LiI(Eu) scintillator as detector. It is
important to mention that the UAZ response matrix
was used to train the ANN contained in NSDANN
[27]. Table 1, shows the counts rate readings of the
UAB BSS system. In table 1, can be seen highlighted
the data used to compare and validate the NSDANN
computer tool.

Figure 3. BBS response matrix of the 3He detector (UAB)

101
0"

100

3"
5"
6"

10-1

[ cm ]

Figure 2. NSDANN computer tool designed with ANN

10"
10-2

12"

10-3

By using NSDANN tool, showed in figure 2, two cases were

carried out in order to probe and to validate the designed tool
comparing the results obtained against classical and
contemporary neutron spectrometry approaches. The first
one case was as follows: a BBS system with a 0.4x0.4 cm2
6
LiI(Eu) scintillator as thermal neutron was used to obtain
the counts rate of a of 252Cf neutron source at 100 on air, in
the Unidad Acadmica de Estudios Nucleares of the
Universidad Autnoma de Zacatecas (UAZ), Mexico. Then
the obtained data were introduced to NSDANN. The results
were compared against the results obtained for the same
count rate but using the neutron spectra unfolding code
known as BUNKI.
For the second case, was used a data set obtained from an
experiment reported and carried out with a new neutron
spectrometric tool (MITOM) proposed by Universidad
Autnoma de Barcelon (UAB), Spain [31]. Several important
differences exist between the UAB experiment by means of
MITOM and the UAZ experiment by means of NSDANN
tool.

10-4

10- 10

10-9

10-8

10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

101

102

Neutron energy [ MeV ]

Figure 4. BBS response matrix of the LiI(Eu) detector

(UAZ)
Table 1. Bonner spheres counts rate of the UAB BSS system
for a realistic spectra used in UAZ-NSDANN
Sphere diameter [inc]
Measured counts rate
2.5
5703e-4
3
6120e-4
4.2
6111e-4
5
5436e-4
6
4452e-4
8
2389e-4
10
1132e-4
12
5120e-4
2.5/Cd
2638e-4
3/Cd
3396e-4
4.2/Cd
4235e-4

3. Results

The most important difference relies in the type of

detector used in each case. In figure 3 can be seen the

191

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:06 from IEEE Xplore. Restrictions apply.

In this work was presented a front end computer

tool, based on the ANN technology and RDANN
methodology, called NDSANN, with the following
characteristics: it is a front end computer tool
application, simple for the use of the end user, fast in
the neutron spectra unfolding process. This tool is
capable to unfold neutron spectra and simultaneously
calculates 13 different equivalent doses in an
independent way, with just the count rates of the BSS
system as entrance data, not requiring the neutron
spectra information to calculate the equivalent doses.
The NSDANN tool after analyzing the input data,
presents the results obtained in both graphical and
numerical way. From de two cases considered in this
work were obtained the following results:
Figure 2 shows the neutron spectra obtained with
NSDANN (blue line) which is compared with the
obtained by means of BUNKI code (red line). Whereas
Figure 6 shows 13 equivalent doses calculated by
means of NSDANN. The equivalent doses are: EAP,
EPA, ERLAT, ELLAT, EROT, EISO, H*(10), Hp,slab(10,0),
Hp,slab(10,30),
Hp,slab(10,45),
Hp,slab(10,15),
Hp,slab(10,60), Hp,slab(10,75). [14]

conditions (neutron detector, Bonner spheres and

response matrix) are too different of the trained ANN
contained in the designed NSDANN computer tool.

Figure 6. Case one 13 equivalent doses calculated by

means of NSDANN (UAZ).

Figure 7. Neutron spectra unfolded by mean of MITOM tool

of UAB for the reference realistic spectra at Canel/T400 de
Cadarache.

Figure 5. Case one - Comparison between the neutron

spectra unfolded by means of BUNKI code and NSDANN at
UAZ

On the other hand, figure 7 shows the neutron

spectra obtained by means of MITOM code at a
nuclear reactor in canel/T400 of Cadarache France,
which is compared with the neutron spectra obtained
by NSDANN, showed in figure 8. Finally in figure 9 is
shown the 13 equivalent doses calculated by
NSDANN. From the results obtained in this work, it is
demonstrated that the designed ANN (contained in the
NSDANN computer tool) has a high generalization
capability, because it is capable to unfold the neutron
spectra, even in the case two, where the spectrometric

Figure 8. Neutron spectra unfolded by mean of UAZSDANN for the reference realistic spectra of case two.

192

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:06 from IEEE Xplore. Restrictions apply.

methodologies, reducing the time spent in the entire

neutron spectra unfolding and dose calculus processes.
One notable characteristic of NSDANN is that for
the equivalent doses calculus is not required the
spectrometric information, because in the ANN
training the neutron spectra and the calculus of the
equivalent doses were considered as independent
problems but they were solved simultaneously just
with one ANN.
From the examples carried out in this work can be
seen that the ANN designed by means of RDANN
methodology and implemented in the NSDANN
computer tool has high performance and generalization
capability, because was capable to generalize in order
to solve a neutron spectra problem with conditions
very different in which the ANN was trained.
Matlab, is a powerful programming software which
has, so far been successfully used for solve ANN
problems in engineering and scientific environments;
however it presents some drawbacks which make
undesirable the use of this programming language
because of several inconveniences. The knowledge
extraction of the synaptic weights as numerical data in
a matrix way, makes possible the design of a
customized graphical user interface outside of the
Matlab environment. Work is being done to this
respect.
Just two experiments were carried out with the
NSDANN computer tool based on ANN technology,
to verify its effectiveness. More experiments are
required in order to verify the results obtained
comparing them with another probed approaches.
The use of ANN technology is a useful alternative
to solve the neutron spectrometry and dosimetry
problems; however, some drawbacks must be solved in
the ANN design process, such as the proper synaptic
weight initialization, and the optimum ANN topology
selection, in order to obtain the best results which this
approach is capable to offer. At present research and
development of tools in this sense are been carried out.
In future works, would be necessary to perform
statistical tests to validate if exists statistical difference
between the desired data and the obtained.

Figure 6. 13 equivalent doses calculated by means of

NSDANN (UAZ) for case two.

From the results obtained in this research, it is

evident that Matlab is a powerful programming
software for ANN design; however it presents some
drawbacks which make undesirable the use of this
programming language in the implementation of the
NSDANN computer tool, because of several
inconveniences such as the NDSANN tool depends
totally on Matlab to be executed, if a user does not
have installed Matlab NSDANN can not be executed.
Another inconvenience is that the front end application
is based in text which makes it not friendly, and
although within Matlab it is possible to develop
graphical interfaces, the first drawback continues being
present. Fortunately Matlab allows extract the
knowledge of the ANN trained and tested, which is
stored at the synaptic weights, as numerical data in a
matrix form, which makes possible the design of a
customized graphical user interface outside of the
Matlab environment. Work is being done to this
respect.

4. Conclusions
In this work, a new computer tool called NSDANN,
based on ANN technology, was designed to unfold
neutron spectra and to simultaneously calculate 13
equivalent doses within Matlab programming
environment. Important characteristics of NSDANN
are the following: first, the user does not need to be a
very expert user; second, the tool is not complex to use
for the end user, third, it is not required a priori
information about the spectra being calculated. The
proposed computer tool reduces significantly the time
required to prepare, to process and to present the
information, in an appropriate way to the researcher, if
compared with the traditional approaches. This tool
overcome some drawbacks of the classical

5. References
[1] R. Begdoni, Neutron spectrometry and dosimetry for
radiation protection around a high energy electrnpositron collider, tesis doctoral, Universidad
Autonoma de Barcelona, 2006
[2] R. Begdoni, C. Domingo, A. Esposito and F. Frenandez,
FRUIT: An operational tool for multisphere neutron
spectrometry in workplaces, Nuclear instruments and
methods in physics research, A 580, 1301-1309 (2007)

193

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:06 from IEEE Xplore. Restrictions apply.

[20] Lippmann R.P., An introduction to computing with

neural networks, IEEE ASSP MAGAZINE, 4(2), 1987.
[21] Kohonen, T., The self-organizing map, Procc. IEEE,
78(9), 1990.
[22] J.M. Ortiz Rodrguez, M.R. Martnez Blanco, and H.R.
Vega carrillo, Problemas crticos en el diseo y
construccin de redes neuronales en la espectrometra
de neutrones, utilizando la metodologa de diseo
robusto de redes neuronales artificiales, 1er. Congreso
panamericano del IRPA, VII congreso regional de
seguridad radiolgica y nuclear, XIV reunion anual de
la SMSR y XVII congreso anual de la SNM, Acapulco,
Guerrero, Mxico (2006).
[23] M.R. Martinez Blanco, J.M. Ortiz Rodrguez, and H.R.
Vega carrillo, Neutron spectrometry and dosimetry by
means of Bonner spheres system and artificial neural
networks applying robust design of artificial neural
networks, 1er. Congreso panamericano del IRPA, VII
congreso regional de seguridad radiolgica y nuclear,
XIV reunin anual de la SMSR y XVII congreso anual
de la SNM, Acapulco, Guerrero, Mxico (2006).
[24] J.M. Ortiz Rodrguez, M.R. Martinez Blanco, and H.R.
Vega carrillo, A proposal to order the neutron data set
in neutron spectrometry using the RDANN
methodology, IX International symposium, XIX
national congress on solid state dosimetry, Mexico city
(2006).
[25] J.M. Ortiz Rodrguez, M.R. Martinez Blanco, and H.R.
Vega carrillo, RDANN a new methodology to solve
the neutron spectra unfolding problem, IX International
symposium, XIX national congress on solid state
dosimetry, Mexico city (2006).
[26] M.R. Martinez Blanco, J.M. Ortiz Rodrguez, and H.R.
Vega carrillo, Neutron spectrometry and dosimetry by
means of Bonner spheres system and artificial neural
networks applying robust design of artificial neural
networks, IX International symposium, XIX national
congress on solid state dosimetry, Mexico city (2006).
[27] J.M. Ortiz Rodrguez, M.R. Martinez Blanco, and H.R.
Vega carrillo, Robust design of artificial neural
networks applying the taguchi methodology and DoE,
CERMA (2006).
[28] Sean C. Miller; AFITBUNKI: A Modified Iterative
Code to Unfold Neutron Spectra from Bonner Sphere
Detector Data, Master's thesis, Air force inst of tech
wright-pattersonafb oh school of engineering, 1993
[29] M. Toms, F. Fernndez, M. Bakali and H. Muller,
(2004), MITOM: A new unfolding code based on a
spectra model method applied to neutron spectrometry,
Radiation protection dosimetry, 110 (1-4), 545-548
(2004).
[30] M. Coeck, V. Lacoste and H. Muller, Neutron
spectrometry around the venus reactor using Monte
Carlo simulations and Bonner spheres measurements,
Radiation protection dosimetry, advance access, 2007.
[31] F. Fernndez, Espectrometra aplicada a la dosimetra,
Radioproteccin, 37, IX, 2002.

[3] V. Lacoste, M. Reginatto, B. Asselineau and H. Muller,

Bonner sphere neutron spectrometry at nuclear
workplaces in the framework of the evidos project,
Radiation protection dosimetry, 125 (1-4), 304-208,
2007
[4] McDonald J.C., Siebert B.R.L., Alberts W.G., Neutron
spectrometry for radiation protection purposes, Nucl.
Instrum. Meth. Phys. Res. A, 476, 2002.
[5] Thomas J. D., Neutron spectrometry for radiation
protection, Radiat. Prot. Dosim., 110, 2004
[6] Matzke M., Unfolding Procedures, Radiation
Protedction Dosimetry, 107, 2003
[7] El Messaoudi M, Chouak A, Lferde M, Cherkaoui R.,
Performance of three different unfolding procedures
connected to Bonner sphere data, Radiat Prot Dosim.,
108, 2004.
[8] Lacoste V., Gressier V., Pochat J.-L., Fernandez F.,
Bakali M., Boassoule T., Characterization of Bonner
spheres systems at monoergetic and thermal neutron
fields, Radiat. Prot. Dosim., 110, 2004.
[9] Vylet V., Response matrix of an extended Bonner
sphere system, Nucl. Instrum. Meth. Phys. Res. A, 476,
2002.
[10] Braga, C.C. and Dias M.S., Application of neural
networks for unfolding neutron spectra measured by
means of Bonner spheres, Nucl. Instrum. Meth. Phys.
Res. A Vol. 476, 2002.
[11] Cordes A., Fehrenbacher G., Schutz R., Sprunk M.,
Hahn K., Hofmann R., Biersack J.P., Wahl W., An
approach to unfold the response of a multi-element
system using an artificial neural network, IEEE Trans.
Nucl. Sci., 45, 1998.
[12] Fehrenbacher G., Schutz R., Hahn K., Sprunk M.,
Cordes, A., Biersack J.P., Wahl W., Proposal of a new
method for neutron dosimetry based on spectral
information obtained by application of artificial neural
networks, Radiat. Prot. Dosim., 83, 1999.
[13] Kardan M.R., Koohi-Fayegh R., Setaycshi S., GhiassiNejad M., Fast neutron spectra determination by
threshold activation detectors using neural network,
Radiat. Measurem. 38, 2004.
[14] H.R. Vega-Carrillo, V.M. Hernandez-Davila, E.
Manzanares-Acua, G.A. Mercado, E. Gallego, E.
Lorente, W.A. Perales, J.A. Robles, Artificial neural
networks in neutron dosimetry,Radiat. Prot. Dosim.,
118 251 (2006).
[15] H.R. Vega-Carrillo, V.M. Hernndez-Dvila, E.
Manzanares-Acua, G. Mercado, M.P. Iiguez, R.
Barquero, F. Palacios, R. Mndez, T. Arteaga, J.M.
Ortz, Neutron spectrometry using artificial neural
networks, Radiat. Measurem., 41 425 (2006).
[16] Fausett L., Fundamentals of neural networks, PrenticeHall, 1993.
[17] Gupta M.M., Lin J., and Homma N., Static and
Dynamic Neural Networks. From Fundamentals to
Advanced Theory, John Wiley & Sons, 2003.
[18] Haykin S., Neural Networks: A Comprehensive
Foundation, Prentice Hall, (1999).
[19] Jain, A.K., Mao J., and Mohuiddin, K.M., Artificial
neural networks, a tutorial, computer, IEEE, 29, 1996.

194

Authorized licensed use limited to: UNIVERSIDAD DE CORDOBA. Downloaded on January 11, 2009 at 15:06 from IEEE Xplore. Restrictions apply.

Captulo 5

Conclusiones
El actual estado del arte de la espectrometra y dosimetra de neutrones por medio de la

tecnologa de IA, como se aprecia en la figura 2.22, es una tecnica muy nueva en este ambito
del conocimiento. A pesar de los resultados alentadores obtenidos, sufre de una serie de incon
venientes que hacen difcil su aplicacion practica
en situaciones experimentales reales, ya que
se observa se carece de herramientas cientfico-tecnologicas que permitan aplicarla de forma

agil
y eficiente.

Lo anterior, como se ilustra en la figura 5.1, puso de manifiesto que la carencia de e stas
bases, as como la falta de herramientas, hacan necesario abrir una lnea de investigacion en
e ste dominio del conocimiento.

De lo expuesto a lo largo del presente documento, al apreciar la figura 5.2, se puede concluir que se alcanzaron los objetivos generales planteados en la seccion 1.4, esto es, el estudio,
y aplicacion de tecnicas de optimizacion, metodologicas y sistematicas,

diseno
en el proceso
o ptimo arquitecturas robustas de RNA capaces de resolver con alto rendimiento y
de diseno
capacidad de generalizacion los problemas relacionados con la espectrometra y dosimetra de
neutrones.

de herComo se observa en la figura 5.2, tambien se cumplio el segundo objetivo: el diseno

y rapida

ramientas tecnologicas que permiten aplicar de forma eficiente, facil

el conocimiento

adquirido por las redes disenadas,

en ambientes laborales, experimentales, y de laboratorio.

197

5. Conclusiones

198

Figura 5.1: Espectrometra de neutrones con el SEEB y tecnicas de IA

Figura 5.2: Contribucion de tecnicas de IA en espectrometra de neutrones con el SEEB

En su mayora, las herramientas disenadas

presentan las siguientes caractersticas: se

crearon bajo el entorno grafico

G de programacion de LabVIEW; operan de forma independi
ente; cuentan con una forma eficiente de entrada y salida de datos; son faciles
de utilizar; los

programas se ejecutan de forma rapida,

5. Conclusiones

199

reconstruccion iterativos, mismos que requieren para la deconvolucion del espectro un usuario
experto, as como informacion a priori (espectro inicial), para la operacion de las herramien

tas disenadas
en este trabajo no es necesaria tal informacion, ya que unicamente
requiere
de las ventajas ya
como datos de entrada, las tasas de conteo provenientes del SEEB, ademas
mencionadas.

de arquitecturas de RNA en el ambito

La estrategia de diseno
de la espectrometra neutronica, por su naturaleza y complejidad, implico que para resolver el problema planteado

fuera necesario dividirlo en multiples

etapas, entre las cuales se pueden destacar las siguientes:
en el proceso de creacion
de las matri1. Pre-procesamiento de la informacion
ces de entrenamiento a ser utilizadas en las etapas de el entrenamiento de las

redes a ser disenadas.

Con el uso de las herramientas disenadas
en esta fase, se consiguio reducir el tiempo general empleado en e sta actividad de meses a horas.

de
2. El empleo de dos metodologas sistematicas
aplicadas en la determinacion

los parametros
optimos
de las redes a ser entrenadas: MDRRNA y RNAE. El
empleo de las herramientas citadas permitio reducir el tiempo empleado en la seleccion

de los parametros
de las arquitecturas de RNA de meses a minutos.
3. El proceso de extraccion del conocimiento de las redes disenadas

y la creacion de un conjunto de codigos para la reconstruccion de espectros de neutrones basado en este conocimiento, disenados

para un SEEB basado en un detector de 6 LiI (Eu) y un SEEB basado en un

detector de 3 He. Las herramientas disenadas

hicieron posible aplicar de forma efectiva,

rapida
y eficiente la tecnologa de RNA en el ambito
de la espectrometra neutronicas,

resolviendo muchos de los problemas asociados con las herramientas clasicas

empleadas
en la actualidad.
y desarrollo de una herramienta de computo

4. El diseno
que permite intercom

parar de forma grafica,

numerica
y estadstica los resultados obtenidos con las

herramientas disenadas
en el punto anterior. Esta
herramienta permitio realizar

una analisis
de los resultados obtenidos, de forma facil,
rapida
y eficiente a traves del

entorno grafico
disenado.

5. Conclusiones

200

y desarrollo de un algoritmo basado en la estadstica de Poisson, im5. El diseno

plementado en las herramientas desarrolladas en el punto anterior, para cuantificar la incertidumbre asociada con las tasas de conteo medidas con los SEEB

utilizados para la espectrometra neutronica.

Las herramientas clasicas
no cuan
tifican las incertidumbres asociadas con el uso del SEEB. Esta
rutina permite cuantificar
el error asociado con las tasas de conteo del SEEB.
6. Aplicar las herramientas desarrolladas en ambientes experimentales reales,
de las

observado su rendimiento y desempeno.

Por las caractersticas de diseno

topologas de RNA disenadas

por otros investigadores, resulta muy complicado aplicar
las en situaciones experimentales reales si no se cuenta con bases solidas en el ambito
de la teora de RNA y en el entorno de ambiente de programacion de Matlab, lo que representa un serio inconveniente. Las herramientas desarrolladas se ajustan al ambiente

experimental por ser autonomas de cualquier programa y ser faciles,

rapidas
y eficientes
en su uso por parte de cualquier persona.

7. Comparar los resultados obtenidos con las herramientas disenadas

contra los

obtenidos con codigos

tradicionales empleados en la actualidad por los distin de las herramientas mencionadas
tos laboratorios. Por las caractersticas de diseno

fue posible realizar multiples

experimentos en la medicion de fuentes de neutrones bajo
distintas condiciones y comparar los resultados contra los obtenidos con las metodologas

clasicas,
observando que las herramientas disenadas
ofrecen ventajas significativas en

cuanto a la facilidad de uso, rendimiento, velocidad de calculo

y manipulacion de informacion.
En vista de que en la actualidad no existen herramientas tecnologicas especficas en el
dominio de la espectrometra de neutrones que permitan aplicar la metodologa de RNAE, en
entrenamiento y prueba de las RNA evolucionadas por medio de AE, se
el proceso de diseno,
utilizo un programa de computo comercial denominado NeuroGenetic Optimizer (NGO). Dicho
programa resulto poco eficiente en el aspecto operativo. Lo anterior sugiere la necesidad de desarrollar una herramienta a la medida para este dominio del conocimiento. En la actualidad
se realizan trabajos tendientes a resolver este inconveniente.

5. Conclusiones

201

de arquitecLa principal y mas seria desventaja en el uso de AE aplicados en el diseno

turas de RNA, es la seleccion de los parametros

de la RNA a ser disenada
y el AE aplicado
para tal fin. En la actualidad se utiliza la tecnica de ensayo y error, que es poco eficiente, consume mucho tiempo y no produce resultados confiables. A la fecha no existen bases cientficas
y metodologicas que permitan resolver e ste gran inconveniente. Sin embargo, a partir de la
experiencia obtenida al crear la metodologa estadstica MDRNA, se observa que combinarla
con la teora de AE, parece ser una buena alternativa en el proceso de solucion de e ste problema. En la actualidad se explora la posibilidad de fusionar las metodologas MDRRAN-AE,

creando una base cientfica y tecnologica hbrida, en la busqueda

de una solucion satisfactoria
a los problemas planteados.

y eficiente. No existe la imperiosa necesidad de proporcionar

se realiza de forma rapida,
facil

al programa informacion a priori o parametro,

a excepcion de las tasas de conteo tomadas a
partir de un SEEB. EL resultado se obtiene en cuestion de segundos, pudiendo almacenarse

los valores calculados para un analisis

posterior y continuar de forma inmediata con la recon

struccion de un numero
ilimitado de espectros de neutrones. Por sus caractersticas de diseno,

e sta herramienta resulta practica

en ambientes experimentales reales.

Se observa que aun

El codigo NSDEann se encuentra en una etapa temprana de diseno.

existen multiples
inconvenientes para aplicarla de forma eficiente, como por ejemplo la depen es comercial, la seleccion de los parametros

dencia con un programa externo que ademas

de
configuracion de los AE, la incapacidad de presentar la informacion obtenida de acuerdo con
las normas establecidas para la espectrometra neutronica, entre otras. Sin embargo, se obser no existen reglas fijas para resolver los problemas antes mencionados, por lo cual
va que aun
es necesario realizar mas investigacion con el objetivo de construir las teoras y herramientas
que permitan resolver los problemas planteados. Lo anterior abre una linea de investigacion
en donde se manifiesta la necesidad de desarrollar herramientas a la medida, con el objeto de

5. Conclusiones

202

realizando trabaaplicar de forma eficiente la tecnologa analizada. En la actualidad se estan

jos tendientes a proponer soluciones a los problemas encontrados.

de las herramientas clasicas

Al comparar el desempeno
con las presentadas en el pre no disposente trabajo de investigacion, se observo que ambas carecen de un elemento comun:
nen de una herramienta que permita cuantificar la incertidumbre asociada con el proceso de
medicion. Este aspecto es importante, ya que la naturaleza propia del proceso de reconstruccion del espectro de neutrones mediante el empleo del SEEB, se ve afectado por varias incertidumbres, entre las cuales se puede mencionar: el proceso de lectura de las tasas de conteo
con el SEEB, el metodo y la calidad con la que fue calculada la matriz de respuesta del sistema, la calidad y eficiencia del detector de neutrones. En vista de la necesidad de cuantificar

la incertidumbre asociada con el proceso de medicion en el ambito

de la espectrometra de neu o un algoritmo de programacion que hace uso de la estadstica de Poisson, para
trones, se disen

cuantificar la incertidumbre asociada a las lecturas tomadas con las esferas del SEEB. Este
se incluyo como una subrutina al codigo NSDann,
algoritmo, en una etapa inicial de diseno
sin embargo, se puede adaptar a codigos semejantes.

En vista de que el SEEB se ve afectado por multiples

incertidumbres, se plantea el uso de
de e stas incertidumbres asociadas en la cuantificacion
la distribucion normal para incluir mas
de un error mas general.

Bibliografa
[1] Fritzsche P. Tools in artificial intelligence. In-Tech, 2008.
[2] Negnevitsky M. Artificial Intelligence, a guide to intelligent systems. Addison Wesley,
2005.
[3] Coppin B. Artificial intelligence iluminated. Jones and Barttlet Publishers, 2004.
[4] Russell S.J. and Norvig P. Inteligencia artificial un enfoque moderno. Prentice Hall,
2004.
[5] Luger G.F. Artificial intelligence structures and strategies for complex problem solving.
Addison-Wesley, 2005.
[6] Luger G.F. and Stubblefield W.A. Artificial Intelligence, structures and strategies for
complex solving problems. Addison Wesley, 1998.
[7] Dorigo M., Bonabeau E., and Theraulaz G. Swarm intelligence from natural to artificial
systems. Oxford University Press, 1999.
[8] Wooldridge M.J. and Veloso M. Artificial Intelligence today: recent trends and developments. Springer, 1998.
[9] Haugeland J. Mind design philosophy psychology artificial intelligence. The MIT Press,
1997.
[10] Yu W. Recent advances in intelligent control systems. Springer-Verlag, 2009.
[11] Munakata T. Fundamentals of the new artificial intelligence, neural, evolutionary, fuzzy
and more. Springer, 2008.

203

BIBLIOGRAFIA

204

[12] Bramer M. and Devedzic V. Artificial intelligence applications and innovations. Kluer
Academic Publishers, 2004.
[13] Tanimoto S.L. The elements of artificial intelligence, an introduction using LISP. Computer Science Press, 1987.
[14] Bratko I. Programming for artificial intelligence. Addison-Wesley, 1986.
[15] Russell S.J. and Norvin P. Artificial intelligence, a moder approach. Prentice Hall, 1995.
[16] Sugumaran V. Distributed artificial intelligence, agent technology, and collaborative applications. Information Science Reference, 2009.
[17] Konar A. Artificial intelligence and soft computing, behavioral and cognitive modelling
of the human brain. CRC Press, 2000.
[18] Weiss G. Multiagent systems a modern approach to distributed modern approach to
artificial intelligence. The MIT Press, 1999.
[19] Partridge D. Artificial intelligence and software engineering, understanding the promise
of the fucture. Glenlake Publishing Company, Ltd., 1998.
[20] Krishnamoorthy C.S. and Rajeev S. Artificial intelligence and expert systems for enginners. CRC Press, 1996.
[21] Lakhmi C. J. Fusion of neural networks, fuzzy sets, and genetic algorithms: industrial
applications. 1st edition, 1998.
[22] Chennakesava R.A. Fuzzy logic and neural networks, basic concepts and applications.
New Age International Publishers, 2008.
[23] Chiong R. Nature-inspired informatics for intelligent applications and knowledge discovery. Information Science Reference, 2009.
[24] Gen M., Green, Katai O., McKay B., Namatame A., Sarker R. A., and Zhang B. T. Intelligent and evolutionary systems. Springer-Verlag, 2009.
[25] Rothlauf F. Representations for Genetic and Evolutionary Algorithms. Springer, 2006.

BIBLIOGRAFIA

205

[26] Whiteson S. and Stone P. Evolutionary function approximation for reinforcement learning. The Journal of Machine Learning Research, 7:877917, month 2006.
[27] Malioukov S.P. Kureichik V.V., Kureichik V.M., and Malioukov A.S. Genetic algorithms
for applied CAD problems. Springer-Verlag, 2009.
[28] Baldi P. and Brunak S. Bioinformatics, the machine learning approach. Mit Press, 2001.
[29] Kasabov N.K. Foundations of neural networs, fuzzy systems, and knowledge engineering.
MIT Press, 1998.
[30] Beale M.H., Hagan M.T., and Demuth H.B. Neural networks toolbox, users guide. Mathworks, 1992. www.mathworks.com/help/pdf_doc/nnet/nnet.pdf.
[31] Fausett L. Fundamentals of neural networks, architectures, algorithms and applications.
Prentice Hall, 1993.
[32] Haykin S. Neural networks: a comprensive foundation. Prentice Hall, 1999.
[33] Jain A.K., Mao J., and Mohiuddin K.M. Artificial neural networks: a tutorial. IEEE:
Computer, 29(3):3144, month 1996.
[34] Zupan J. Introduction to artificial neural network methods: what they are and how to
use them. Acta Chimica Slovenica, 41(3):327352, month 1994.
[35] Taylor J.G. Mathematical approaches to neural networks. North-Holland Mathematical
library, 1993.
[36] Gelatt C.D., Kirkpatrick S., and Vecchi M.P. Evolutionary artificial neural network for
selecting flexible manufacturing systems under disparate level-of-satisfaccion of decision maker. Science, 220(4598):671680, 1983.
[37] Lippmann R. An introduction to computing with neural nets. IEEE ASSP Magazine,
4(2):422, 1987.
[38] Arbib M.A. Brain theory and neural networks. The Mit Press, 2003.
[39] Perlovsky L.I. Neural networks and intellect: using model-based concepts. Oxford University Press, 2001.

BIBLIOGRAFIA

206

[40] Sivanandam S.N. and Deepa S.N. Introduction to genetic algorithms. Springer-Verlag,
2008.
[41] Rothlauf F. Representations for Genetic and Evolutionary Algorithms. Springer, 2006.
[42] Yao X. Evolutionary artificial neural networks. International Journal of Neural Systems,
4(3):203222, month 1993.
[43] Affenzeller M., Winkler S., Wagner S., and Beham A. Genetic algotithms and genetic
programming, modern concepts and practical applications. CRC Press, 2009.
[44] Haupt R. L. and Haupt S. E. Practical genetic algorithms. Wiley, 2004.
[45] Mitchell M. An introduction to genetic algorithms. MIT Press, 1998.
[46] Periaux J. and Winter G. Genetic algorithms in engineering and computer science. John
Wiley & Sons, 1995.
[47] Goldberg D.E. Genetic Algorithms in search, optimization, and machine learning. Addison Wesley, 1989.
[48] Bhattacharya A. Evolutionary artificial neural network for selecting flexible manufacturing systems under disparate level-of-satisfaccion of decision maker. International
Journal of Innovative Computing, Information and Control, 3(1):131140, 2007.
[49] Melin P. and Castillo O. Hybrid intelligent system for pattern recognition using soft
computing, an evolutionary approach for neural networks and fuzzy systems. Springer,
2005.
[50] Rasskin-Gutman D. Chess metaphors, artificial intelligence and the human brain. The
MIT Press, 2009.
[51] Floreano F. and Mattiussi C. Bio-inspired artificial intelligence, theories, methods and
technologies. The MIT Press, 2008.
[52] Periaux J. and Winter G. Genetic algorithms in engineering and computer science. John
Wiley & Sons, 1995.
[53] Mitchell M. An introduction to genetic algorithms. MIT Press, 1998.

BIBLIOGRAFIA

207

R., Hahn K., Sprunck M.E., Biersack J.P., and Wahl W. Pro[54] Fehrenbacher G., Schutz
posal of a new method for neutron dosimetry based on spectral information obtained
by application of artificial neural networks. Radiation Protection Dosimetry, 83(4):293
301, 1999.
[55] Cordes E., Fehrenbacher G., Schutz R., Sprunck M., Hahn K., Hofmann R., Biersack J.
P., , and Wahl W. An approach to unfold the response of a multi-element system using
an artificial neural network. IEEE Transactions on Nuclear Science, 45(3):14641469,
1998.
[56] Dreyfus G. Neural networks, methodology and applications. Springer, 2005.
[57] Bassis S., Apolloni B., and Marinaro M. New directions in neural networks. IOS Press,
2009.
[58] Graupe D. Principles of artificial neural networks. World Scientific, 2007.
[59] Galushkin A.I. Neural networks theory. Springer, 2007.
[60] Lakhmi J. and Fanelli A.M. Recent advances in artificial neural networks design and
applications. CRC Press, 2000.
[61] Freeman J.A. and Skapura D.M. Neural networks: algorithms, applications, and programming techniques. Addison Wesley Longman Publishing Co., Inc., 1991.
[62] Braga C.C. and Dias M.S. Application of neural networks for unfilding neutron spectra
measured by means of Bonner spheres. Nuclear Instruments and Methods in Physics
Research Section A, 476(1-2):252255, 2002.
[63] Kardan M.R., Setayeshi S., Koohi-Fayegh R. and Ghiassi-Nejad M. Neutron spectra
unfolding in Bonner spheres spectrometry using neural networks. Radiation Protection
Dosimetry, 104(1):2730, 2003.
[64] Kardan M.R., Koohi-Fayegh R., Setayeshi S., and Ghiassi-Nejad M. Neutron spectra
unfolding in Bonner spheres spectrometry using neural networks. Radiation Protection
Dosimetry, 104(1):2730, 2004.

BIBLIOGRAFIA

208

[65] Kardan M. R., Koohi-Fayegh R., Setayeshi S., and Ghiassi-Nejad M. Fast neutron spectra determination by threshold activation detectors using neural networks. Radiation
Measurements, 38:185191, 2004.
[66] Lewis E.E. Fundamentals of nuclear reactor physics. Academic Press, 2008.
[67] Jevremovic T. Nuclear principles in engineering. Springer, 2005.
[68] Rich J., Basdevant J.L., and Spiro M. Fundamentals in nuclear physics. Springer, 2005.
[69] Attix F.H. Introduction to radiological physics and radiation dosimetry. Wiley-VCH,
2004.
[70] Lilley J. Nuclear physics, principles and applications. John Wiley & Sons, Ltd., 2001.
[71] Bromley D.A. Detectors in nuclear science. Nuclear instruments and methods, 162,
1979.
[72] Roy R.R. and Nigam B.P. Nuclear physics, theory and experiment. John Wiley & Sons,
Inc., 1967.
[73] Shani G. Radiation dosimetry, instrumentation and methods. CRC Press, 2001.

[74] Bedogni R., Domingo C., Esposito A., and Fernandez

F. FRUIT: An operational tool for
multisphere neutron spectrometry in workplaces. Nuclear Instruments and Methods in
Physics Research A, 580(3):13011309, 2007.
[75] Leroy C. and Rancoita P.G. Principles of radiation interaction in matter and detection.
World scientific, 2009.
[76] Murray R.L. Nuclear energy, an introduction to the concepts, systems, and applications
of nuclear processes. World Scientific, 2000.
[77] Heyde K. Basic ideas and concepts in nuclear physics. Institute of Physics IOP Publishing Ltd., 1999.
[78] Carron N.J. An introduction to the passage of energetic particles trough matter. Taylor
& Francis, 2007.
[79] Meyerhof W.E. Elements of nuclear physics. McGraw Hill, 1967.

BIBLIOGRAFIA

209

[80] Ramayya A.V., Hamiiton J.H., and Carter H.K. Fission and properties of neutron-rich
nuclei. World scientific Publishing Co. Re. Ltd., 2008.
[81] Lim Y.K. Problems and solutions on atomic, nuclear and particle physics. World Scientific, 2000.
[82] Knoll G.F. Radiation detection and measurement. John Wiley & Sons, 3a edition, 1999.
[83] Watt D.E. Quantities for dosimetry of ionizing radiations in liquid water. Taylor &
Francis, 1996.
[84] Cember H. Introduction to health physics. McGraw-Hill, 3a edition, 1996.
[85] IAEA. Compendium of neutron spectra in criticality accident dosimetry. Technical Report 180, 1979.
[86] IAEA. Compendium of neutron spectra and detector responses for radiation protection
purposes. Technical Report 318, 1990.
[87] Lamarsh J.R. Introduction to nuclear engineering. Addison-Wesley Publishing Company, 1983.
[88] Krane K.S. Introductory nuclear physics. Jhon Wiley & Sons, Inc., 1983.
[89] Bramblett R.L., Ewing R.I., and Bonner T.W. A new type of neutron spectrometer. Nuclear Instruments and Methods, 9:112, 1960.
[90] Bonner T.W. Measurements of neutron spectra from fission. Nuclear Physics, 23:116
121, 1961.
[91] Axton E.J. The effective centre of a moderating sphere used as instrument for fast
neutron flux measurement. Journal of Nuclear Energy, 26:581573, 1972.
[92] Thomas D.J.

Neutron spectrometry for radiation protection.

Radiation Protection

Dosimetry, 110(1-4):141149, 2004.

[93] Alevra A.V., Cosack M., Hunt J.B., Thomas D.J., and Schraube H. Experimental determination of the response of four Bonner sphere sets to monoenergetic neutrons (II).
Radiation Protection Dosimetry, 40(2):91102, 1992.

BIBLIOGRAFIA

210

[94] Matzke M. Unfolding procedures. Radiation Protection Dosimetry, 107(1-3):155174,

2003.
M., Fernandez

[95] Tomas
F., Bakali M., and Muller H. MITOM: a new unfolding code
based on a spectra model method applied to neutron spectrometry. Radiation Protection Dosimetry, 110(1-4):545548, 2004.
[96] Lacoste V., Reginatto M., Asselineau B., and Muller H. Bonner sphere neutron spectrometry at nuclear workplaces in the framework of the evidos project. Radiation Protection
Dosimetry, 125(1-4):304308, 2007.

E., Mercado Sanchez

[97] Vega-Carrillo H.R., Hernandez-D

avila
V.M., Manzanares-Acuna
W.A., and Robles-Rodrguez J.A. Artificial
G.A., Gallego E., Lorente A., Perales-Munoz
neural networks in neutron dosimetry. Radiation Protection Dosimetry, 118(3):251259,
2006.

E., Mercado-Sanchez

[98] Vega-Carrillo H.R., Hernandez-D

avila
V.M., Manzanares-Acuna

R.,
G.A., Iniguez
de la Torre M.P., Barquero R., Preciado-Flores S., Mendez-Villafane
Arteaga-Arteaga T., and Ortiz-Rodrguez J.M. Neutron spectrometry using artificial
neural networks. Radiation Measurements, 41:425431, month 2006.

E., Gallego E., Lorente

[99] Vega-Carrillo H.R., Hernandez-D
avila
V.M., Manzanares-Acuna

A., and Iniguez

M.P. Artificial neural networks technology for neutron spectrometry and
dosimetry. Radiation Protection Dosimetry, 126(1-4):408412, 2007.

[100] Vega-Carrillo H.R., Ortiz-Rodrguez J.M., Hernandez-D

avila
V.M., Martnez-Blanco

M.R., Hernandez-Almaraz
B., Ortiz-Hernandez
A., and Mercado-Sanchez
G.A. Different spectra with the same neutron source. Revista Mexicana de Fsica S, 56(1):3539,
month 2009b.

[101] Vega-Carrillo H.R., Martnez-Blanco M.R., Hernandez-D

avila
V.M., and OrtizRodrguez J.M. Ann in spectroscopy and neutron dosimetry. Journal of Radioanalytical
and Nuclear Chemistry, 281(3):615618, month 2009a.

E., Gallego E., Lorente

[102] Vega-Carrillo H.R., Hernandez-D
avila
V.M., Manzanares-Acuna

A., and Iniguez

M.P. Artificial neural networks technology for neutron spectrometry and
dosimetry. Radiation Protection Dosimetry, 126(1-4):408412, month 2007b.

BIBLIOGRAFIA

211

[103] Vega-Carrillo H.R., Martnez-Blanco M.R., Hernandez-D

avila
V.M., and OrtizRodrguez J.M. Spectra and dose with ANN of 252Cf, 241Am-Be, and 239Pu-Be. Journal
of Radioanalytical and Nuclear Chemistry, 281(3):615618, 2009.

[104] Vega-Carrillo H.R., Martnez-Blanco M.R., Hernandez-D

avila
V.M., and OrtizRodrguez J.M. Spectra and dose with ANN of 252Cf, 241AmBe and 239PuBe. Journal
of Radianalitycal and Nuclear Chemistry, 281(3):615618, 2009.
[105] Kishan M., Chilukuri K., and Sanjay R. Elements of artificial neural networks. The MIT
Press, 2000.
R., Pazos A., Rivero D., and Pereira J. Artificial neural network
[106] Dorado J., Rabunal
development by means of genetic programming with graph codification. Proceedings of
world academy of science, engineering and technology, 15:209214, 2006.
[107] Leardi R. Nature-inspired methods in chemometrics: genetic algorithms and artificial
neural networks. Elsevier, 2003.
[108] Hammer B. and Villmann T. Mathematical aspects of neural networks. pages 5972,
2003.
[109] Zupan J. Introduction to artificial neural network (ANN) methods: what they are and
how to use them. Acta Chimica Slovenica, 41(3):327352, 1994.
[110] Fausett L.V. Fundamentals of neural networks. Prentice Hall, 1 edition, 1993.
[111] Artificial neural networks matlab tool box, ver.7.0, 2007.
[112] Mohan K., Mehrotra C.K., and Ranka S. Elements of artificial neural networks. -, 1996.
[113] Gupta M.M., Jin L., and Homma N. Static and dynamic neural networks: from fundamentals to advanced theory. 2003.
[114] Haykin S. Neural networks: a comprehensive foundation. Prentice-Hall, 2a edition, 1999.
[115] Frenie J.A. and Jiju A. Teaching the taguchi method to industrial engineers. MCB
University Press, 50(4):141149, 2001.

BIBLIOGRAFIA

212

[116] Tam S.C., Chen W.L., Chen Y. H., and Zheng H.Y. Application of taguchi method in the
optimization of laser micro-engraving of photomasks. International Journal of Materials
& Product Technology, 11(3-4):333344, 1996.
[117] Peterson G.E., St. Clair D.C., Aylward S.R., and Bond W. E. Using taguchis method
of experimental design to control errors in layered perceptrons. IEEE Transactions on
Neural Networks, 6(4):949961, month 1995.
[118] Packianather M.S. and Drake P.R. Modelling neural network performance through response surface methodology for classifying wood veneer defects. Proceedings of the Institution of Mechanical Engineers, Part B, 218(4):459466, month 2004.
[119] Packianather M.S., Drake P.R., and Rowlands H. Optimizing the parameters of multilayered feedforward neural networks through taguchi design of experiments. Quality
and Reliability Engineering International, 16:461473, month 2000.
[120] Tseng C.H. and Lin T.Y. Optimum design for artifcial neural networks: an example in
a bicycle derailleur system. Engineering Applications of Artificial Inteligence, 13:314,
2000.
[121] Alavala C.R. Fuzzy logic and neural networks basic concepts & applications. New Age
International Publishers, 1996.
[122] Zenkouar K., El Fadili H., and Qjidaa H. Evolving neural networks using moment
method for handwritten digit recognition. Proceedings of world academy of science, engineering and technology, 6:304307, 2005.
[123] Stone P. and Whiteson S. Evolutionary function approximation for reinforcement learning. The Journal of Machine Learning Research, 7:877917, 2006.
[124] Rizki M.M.and Wicker D. and Tamburino L.A. E-net: Evolutionary neural networks
synthesis. Neurocomputing, 42:171196, 2002.
[125] Delgado M., Blanco A., and Pegalajar M.C. A real-coded genetic algorithms for training
recurrent neural networks. Neural networks, 14:93105, 2001.
[126] Gen M. and Cheng R. Genetic algorithms engineering optimization. John Wiley & Sons,
Inc., 2000.

BIBLIOGRAFIA

213

[127] Goldberg D.E. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison Wesley, 1989.
[128] Reeves C.R. and Rowe J.E. Genetic algorithms principles and perspectives, a guide to
GA theory. Kluwer Academic Publishers, 2002.
[129] Chambers D.L. Handbook of genetic algorithms. New frontiers, volume 2. CRC Press,
1995.
[130] Chambers D.L. Handbook of genetic algorithms. complex coding systems, volume 3. CRC
Press, 1999.
[131] Chambers D.L. Handbook of genetic algorithms. Applications, volume 1. CRC Press,
2001.
[132] De Jong R.P.and Kenneth A. and ROWE J.E. Foundations of genetic algotithms. CRC
Press falta este dato, 2000.
[133] De Jong K.A. Evolutionary computation, an unified approach. The MIT Press, 2006.
[134] Glover F. and Kochenberger G.A. Handbook of meteheuristics. Kluer Academic Publishers, 2003.
[135] Chong E.K.P. and Zak S.H. An introduction to optimization. John Wiley & Sons, Inc.,
2001.
[136] Lee K.Y. and El-Sharkawi A. Modern heuristic optimization techniques, theory and applications to power systems. John Wiley & Sons, Inc., 2008.
[137] Klapdor-Kleingrothaus H.V. and Staudt A. Non-accelerator particle physics. IOP Publishing Ltd., 1995.
[138] Asimov I. Understanding physics, the electron, proton and neutron. New American
Library, 1966.
[139] Gutberlet T., Fitter J., and Katsaras J. Neutron scattering in bilogy, techniques and
applications. Springer, 2006.

BIBLIOGRAFIA

214

[140] Siebert B. R.L., McDonald J.C., and Alberts W.G. Neutron spectrometry for radiation
protection purposes. Nuclear Instruments and Methods in Physics Research A, 476(12):347352, 2002.
[141] Wielunski M. Intercomparison excercise whith mev neutrons using various electronic
personal dosimeters. Radiation Measurements, 43:10631067, 2008.
[142] Cummings F. Neutron measurements using the universal radiation spectrum analyzer.
Radiation Protection Management, 22(2):19, 2005.

[143] Fernandez
F. Espectrometra de neutrones aplicada a la dosimetra. Radioproteccion,
IX(34):3644, 2002.
[144] Dearnaley G. Fast neutron physics, page 173. University of Chicago, 1963.
[145] Reuss P. Neutron physics. EDP Sciences, 2008.
[146] Brooks F. D. and Klein H. Neutron spectrometry, historical review and present status.
Nuclear Instruments and Methods in Physics Research A, 476:111, 2002.
[147] McDonald J. C., Siebert B.R.L., and Alberts W.G. Neutron spectrometry for radiation
protection purposes. Nuclear Instruments and Methods in Physics Research A, 476(12):347352, month 2002.
[148] Tripathy S. P. and Tommasino L. Hardness ratios of different neutron spectra. Radiation Protection Dosimetry, 110(1-4):227231, 2004.
[149] Sims C.S. and Dickson H.W. Neutron dosimetry intercomparison studies. Radiation
Protection Dosimetry, 10(1-4):331340, 1985.
[150] Coyne J.J: Gertenberg H.M., Caswell R.S., and Axton E.J. Basic data necessary for
neutron dosimetry. Radiation Protection, 23(1-4):1117, 1988.
[151] Peurrung A.J. Recents developments in neutron detection. Nuclear Instruments and
Methods in Physics Research Section A, 443:400415, 2000.
[152] Seghour A. and Sens J.C. Neutron spectrometry at strasbourg university reactor. Nuclear Instruments and Methods in Physics Research Section A, 420:243248, 1999.

BIBLIOGRAFIA

215

[153] Klein H. and Neumann S. Neutron and photon spectrometry with liquid scintillation
detectors in mixed fields. Nuclear Instruments and Methods in Physics Research Section
A, 476:132142, 2002.
[154] Harvey J.R., Harrison K.G, and Boot S.J. The calibration of neutron instruments and
dosimeters at intermediates energies. Nuclear Instruments and Methods, 148:511520,
1978.
[155] Ionita I. Focusing conditions in neutron spectrometry. Nuclear Instruments and Methods
in Physics Research Section A, 498:313336, 2002.
[156] Harvey J.A. and Hill N.W. Scintillation detectors for neutron physics research. Nuclear
Instruments and Methods, 162:507529, 1979.
[157] Marion J.B. and Fowler J.F. Fast neutron physics, volume 1 & 2. Interscience, New York,
1960.
[158] Delgado A. Mixed field dosimetry., 2000.
[159] Grosshoeg G.

Neutron ionization chambers.

Nuclear Instruments and Methods,

162:125160, 1979.
[160] Hertel E.N. and Davidson J.W. The response of bonner spheres to neutrons from thermal
energies to 17.3 mev. Nuclear Instruments and Methods in Physics Research Section A,
238:509516, 1985.
[161] Rickard J.C. Neutron spectrometry in the energy range 1-100kev using the 6li sandwich
spectrometer. Nuclear Instruments and Methods, 82:287288, 1970.
[162] Barnhart T.E., Haney J.H., and Zaidins C.S. Extraction of neutron spectral information
from bonner-sphere data. Nuclear Instruments and Methods in Physics Research Section
A, 431:551555, 1999.
[163] Thomas D.J. and Alevra A.V. Bonner sphere spectrometers- a critical review. Nuclear
Instruments and Methods in Physics Research Section A, 476:1220, 2002.
[164] Bedogni R. Lembo L., Esposito A., and Morelli M. Determination of the neutron spectra
around an 18 mv medical linac with a passive bonner sphere spectrometer based on gold
foils and tld pairs. Radiation Measurements, 43:10381043, 2008.

BIBLIOGRAFIA

216

[165] Mazrou H. Characterization of tue crna bonner spectrometer based on 6lii scintillator
exposed to an 241ambe. Radiation Measurements, 43:10951099, 2008.
[166] Souto E. B. Development of a brazilian-neutron dosimeter. Nuclear Instruments and
Methods in Physics Research A, 226:31743177, 2008.
[167] Mares V., Garny S., and Ruhm W. Response functions of a bonner sphere spectrometer
calculated with geant4. Nuclear Instruments and Methods in Physics Research Section
A, 604:612617, 2009.
[168] Goldhagen P. Bonner-sphere neutron spectrometry, 1997.
k M. Mesurements of h*(10) in reference neutron
[169] Jozefowicz K., Golink N., and Kral
fields using bonner sphere spectrometry and let spectrometry. Nuclear Instruments and
Methods in Physics Research Section A, 476:365368, 2002.

[170] Lacoste V., Gressier V., Pochat J.L., Fernandez

F., Bakali M., and Bouassoule T. Characterization of bonner sphere systems at monoenergetic and thermal neutron fields.
Radiation Protection Dosimetry, 110(1-4):529532, month 2004.
[171] Vylet V. Response matrix of an extended bonner sphere system. Nuclear Instruments
and Methods in Physics Research A, 476:2630, month 2002.
[172] Miller W.H. and Brugger R.M. Additional experimental tests of the bonner sphere neutron spectrometer. Nuclear Instruments and Methods in Physics Research Section A,
236:333339, 1985.
[173] Alevra A.V and Plostinaru V.D. Characterization of the ipne bonner sphere spectrometer by comparison with the ptb system. Nuclear Instruments and Methods in Physics
Research, A(476):2125, 2002.
[174] Wiegel B. and Alevra A.V. Nemus- the ptb neutron multisphere spectrometer: Bonner
spheres and more. Nuclear Instruments and Methods in Physics Research Section A,
476:3641, 2002.
[175] Vega-Carrillo H.R., Wehring B.W., Veinot K.G., and Hertel N.E. Response matrix for
a multisphere spectrometer using a 6LiF thermoluminescence dosimeter. Radiation
Protection Dosimetry, 81(2):133139, 1999.

BIBLIOGRAFIA

217

E., Mercado-Sanchez

[176] Vega-Carrillo H.R., Hernandez-D

avila
V.M., Manzanares-Acuna

R., ArteagaG.A., Iniguez

de la Torre M.P., Barquero R., Palacios F., Mendez-Villafane
Arteaga T., and Ortiz-Rodrguez J. M. Neutron spectrometry using artificial neural
networks. Radiation Measurements, 41:425431, 2006.
[177] Su C. and Sarkar T.K. A multiscale moment method for solving fredholm integral equation of the first kind. Progress In Electromagnetics Research, 17:237264, 1997.
[178] Dietz E., Weyrauch M., and Matzke M. Determination of neutron spectra using the
programs gnsr and spectrix. Nuclear Instruments and Methods in Physics Research
Section A, 476:208212, 2002.
[179] Joyce M.J., Peyton A.J., Weaver J.A., and Roskell J. Recent improvements to a transportable neutron spectrometer (tns). Nuclear Instruments and Methods in Physics Research A, 476(1-2):143148, 2002.
[180] Alevra A. V. and Wiegel B. Nemusfthe ptb neutron multisphere spectrometer: Bonner
spheres and more. Nuclear Instruments and Methods in Physics Research A, 476:3641,
2002.
[181] Agosteo S., Bedogni R., Caresana M., Esposito A., Fehrenbacher G., Ferranini M., Hphmann E., Hranitzki C., Kasper A., Khurana S., Mares V., Reginatto M., Rollet S., Ruhm
W., Schardt D., Silari M., Simmer G., Wiegel B., and Weitzenegger E. Intercomparison
of radiation protection devices in high-energy stray neutron field, part ii: Bonner sphere
spectrometry. Radiation Measurements, 44:660672, 2009.
[182] Thomas D.J.

Neutron spectrometry for radiation protection.

Radiation Protection

Dosimetry, 110(1-4):141149, month 2004.

[183] M. Wielunski, W. Wahl, N. EL-Faramawy, W. Ruhm,

M. Luszik-Bhadra, and H. Roos.
Development of a brazilian gamma-neutron dosimeter. Nuclear Instruments and Methods in Physics Research B, 266(12-13):31743177, month 2008.

[184] Wielunski M., Wahl W., EL-Faramawy N., Ruhm

W., Luszik-Bhadra M., and Roos H.
Intercomparison exercise whith mev neutrons using various electronic personal dosimeters. Radiation Measurements, 43(2-6):10631067, month 2008.

BIBLIOGRAFIA

218

[185] Mazrou H., Sidahmed T., Idiri Z., Lounis-Mokrani Z., Bedek Z., and Allab M. Characterization of the crna bonner sphere spectrometer based on 6lii scintillator exposed to
an 241ambe neutron source. Radiation Measurements, 43:10951099, month 2008.
[186] Lacoste V., Reginatto M., Asselineau B., and Muller H. Bonner sphere neutron spectrometry at nuclear workplaces in the framework of the evidos project. Radiation Protection
Dosimetry, 125(1-4):304308, month 2007.

[187] Bedogni R., Domingo C., Esposito A., and Fernandez

F. Fruit: an operational tool for
multisphere neutron spectrometry in workplaces. Nuclear Instruments and Methods in
Physics Research A, 580:13011309, month 2007.
[188] El Messaoudi M., Chouak A., Lferde M., and Cherkaoui R. Performance of three different unfolding procedures connected to bonner sphere data. Radiation Protection Dosimetry, 108(3):247253, month 2004.
[189] Miller S.C. Afitbunki: a modified iterative code to unfold neutron spectra from Bonner
sphere detector data. Masters thesis, 1993.
[190] weezy J., Hertel N.E., and Veinot K. BUMSBonner sphere unfolding made simple: an
html based multisphere neutron spectrometer unfolding package. Nuclear Instruments
and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and
Associated Equipment, 476(1-2):263269, 2002.
[191] Bedogni R. Neutron spectrometry and dosimetry for radiation protection around a high
energy-positron colitter. PhD thesis, 2006.
[192] Miller W.H. and Meyer W. Standarization of neutron spectrum unfolding codes: tests
using idealized, calculated response functions.

Nuclear Instruments and Methods,

205:185195, 1983.
[193] Vylet V. Response matrixof an extended bonner sphere system. Nuclear Instruments
and Methods in Physics Research A, 479:2630, 2002.
[194] Mares V. and Schraube H. Improved response matrices of bonner sphere spectrometers
with 6lii scintillation detector and 3he proportional counter between 15 and 100 mev.
Nuclear Instruments and Methods in Physics Research Section A, 366:203206, 1995.

BIBLIOGRAFIA

219

[195] Mares V. and Schraube H. Evaluation of the response matrix of a bonner sphere spectrometer with lii detector form thermal energy to 100 mev. Nuclear Instruments and
Methods in Physics Research Section A, 337:461473, 1994.
[196] Bakshi A.K. Sathian V. Tripathy S.M. Vega-Carrillo H.R. Nandy M. Sarkar P.K., Tripathy S.P., and Sharma D.N. Measurement of 241ambe spectra (bare and pb-covered)
using tld pairs in multi-spheres: spectrum unfolding by differents methods. Nuclear
Instruments and Methods in Physics Research Section A, 598:556560, 2009.
[197] Sunil C. Nandy M. Sarkar P.K. Sharma D.N., Tripathy S.P., and Mukherjee B. Activation foils unfolding for neutron spectrometry: Comparison of different deconvolution
methods. Nuclear Instruments and Methods in Physics Research Section A, 583:421
425, 2007.
[198] Siebert B.R.L. Assesment of sensitivities and uncertainties in monte carlo particle
transport calculations for neutron spectrometry. Nuclear Instruments and Methods in
Physics Research Section A, 476:256262, 2002.
[199] Hertel N., Sweezy J., and Veinot K. Bums- bonner sphere unfolding made simple: an
html based multisphere neutron spectrometer unfolding package. Nuclear Instruments
and Methods in Physics Research Section A, 476:263269, 2002.
[200] Itoh S. A fundamental study of neutron spectra unfolding based on the maximum
likelihood method. Nuclear Instruments and Methods in Physics Research Section A,
251:144155, 1986.
[201] Lacoste V., Coeck M., and Muller H. Neutron spectrometry around the venus reactor
using monte carlo simulations and bonner spheres measurements. Radiation Protection
Dosimetry, 0(2007), 2007.
[202] Chambless D.A. and Broadway J.A. Comments on neutron spectral unfolding using the
monte carlo method. Nuclear Instruments and Methods, 214:543545, 1983.
[203] OBrien K. and Sanna R. neutron spectral unfolding using the monte carlo method.
Nuclear Instruments and Methods, 185:277286, 1981.

BIBLIOGRAFIA

220

[204] Matzke M. and Weise K. Neutron spectrum unfolding by the monte carlo method. Nuclear Instruments and Methods in Physics Research Section A, 234:324330, 1985.
[205] Fernandez F. Tomas M. Bakali M. Castelo J. Carmena P., Bouassoule T., and Lara L.
Monte carlo calculations and experimental calibrations of bonner sphere systems with
a new cylindrical helium-3 proportional counter. Radiation Measurements, 34:199202,
2001.
[206] Chouak A. Lferde M., El Messaoudi M., and Cherkaoui R. Performance of three different
unfolding procedures connected to bonner sphere data. Radiation Protection Dosimetry,
108(3):247253, 2004.
[207] Routti J.T and Sandberg J.V. General purpose unfolding program loui78 with linear and
non linear regularizations. Computer Physics Communications, 21:119144, 1980.
[208] Matzke M. Estimation of dose equivalent from reactions rates of bonner spheres without
using a priori fluence information. Radiation Protection Dosimetry, 23(1-4):297300,
1988.
[209] Mares V. Weitzenegger E., Simmer G., and Ruhm W. Iterative unfolding for bonner
sphere spectrometers using the msandb code - sensitivity analysis and dose calculation.
Radiation Measurements, in press:19, 2009.
[210] Roberts N.J. Investigation of combined unfolding of neutron spectra using the umg
unfolding codes. Radiation Protection Dosimetry, 126(1-4):398403, month 2007.
[211] Reginatto M. and Zimbal A. Bayesian and maximum entropy methods for fusion diagnostic measurements with compact neutron spectrometers. Review of Scientific Instruments, 79(2):398403, month 2008.
[212] Freeman D.W., Edwards D.R., and Bolon A.E. Genetic algorithms - a new techniqe for
solving a neutron spectrum unfolding problem. Nuclear Instruments and Methods in
Physics Research A, 425, 1999.
[213] Mukherjee B. Andi-03: a genetic algorithm tool for the analysis of activation detector
data to unfold high-energy neutron spectra. Radiation Protection Dosimetry, 110(14):249254, month 2004.

BIBLIOGRAFIA

221

[214] Mukherjee B. A high-resolution neutron spectra unfolding method using the genetic
algorithm technique. Nuclear Instruments and Methods in Physics Research A, 479(12):247251, 2002.
[215] Green R., Crout N.M.J., Taylor G.C., Koohi-Fayegh R., and Scott M.C. Neural network
unfolding of photon and neutron spectra using an ne-213 scintillation detector. Nuclear
Instruments and Methods in Physics Research Section A, 329:269276, 1993.
[216] Sara A.P., Senada A., and Vladimir P. Detector response unfolding using artificial neural
networks. Nuclear Instruments and Methods in Physics Research Section A, 565:742
752, 2006.
[217] Bonyadi M.R., Etaati G.R., Sharghi A., and Shahriari M. Unfolding the neutron spectrum of a ne213 scintillator using artificial neural networks. Applied Radiation and
Isotopes, 67:19121918, 2009.
[218] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. UserFillol A., Mendez-Villafane
friendly interface for run MCNP code for radiation physics. In VI International Symposium on Radiation Physics. Zacatecas, Zac., Mexico, 7-10/Mar/2010.
[219] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Robust design of
artificial neural networks applying the Taguchi methodology and DoE. Proceedings
of the Electronics, Robotics and Automotive Mechanics Conference (CERMA06), IEEE
Computer Society, 2006.
[220] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. NeuFillol A., Mendez-Villafane
tron spectrometry using artificial neutral networks for a Bonner sphere spectrometer
with a He-3 detector. Revista Mexicana de Fsica S, 57(1):6971, 2011.
[221] Patrick Hui C.L. Artificial Neural Networks - Applications. INTECH, 2011.
[222] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Evolutive artificial neural networks for neutron spectra unfolding. Congreso de la Sociedad
Nuclear Americana, Nov/2009.

BIBLIOGRAFIA

222

[223] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Artificial neural networks modeling evolved genetically, a new approach applied in neutron
spectrometry and dosimetry research areas. Proceedings of the Electronics, Robotics and
Automotive Mechanics Conference (CERMA08), IEEE Computer Society, pages 387392,
2008.
[224] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. PerFillol A., Mendez-Villafane
formance of artificial neutral networks and genetical evolved artificial neutral networks
unfolding techniques. Revista Mexicana de Fsica S, 57(1):8992, 2011.
[225] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Redes
neuronales evolutivas: una nueva alternativa para la espectrometra de neutrones. In
XI International Symposium / XXI National Congress on Solid State Dosimetry. Hermosillo, Son., Mexico, 28-2/Oct/2009.
[226] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Neutron spectrometry and dosimetry based on a new approach called genetic artificial neural networks. In 12th Congress of the International Radiation Protection Association
(IRPA12), pages xxyy, International Atomic Energy Agency, Vienna, Oct 2010. IAEA
Procceding Series STI/PUB/1460 1-9.
[227] Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann, a LabVIEW tool for neutron spectrometry and dosimetry based on the RDANN methodology.
Proceedings of the Electronics, Robotics and Automotive Mechanics Conference (CERMA09), IEEE Computer Society, 00:131136, 2009.
[228] Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann: Espectrometra y dosimetra de neutrones. Mexico Nuclear, 9(2):1929, 2009.
[229] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Una
comparacion en la reconstruccion de espectros de neutrones utilizando tecnicas iter
ativas clasicas
y las basadas en inteligencia artificial.

In XI International Sympo-

sium / XXI National Congress on Solid State Dosimetry. Hermosillo, Son., Mexico, 282/Oct/2009.

BIBLIOGRAFIA

223

[230] Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Espectrometra

y dosimetra de neutrones en aceleradores de partculas de uso medico, utilizando NSDANN. In XI International Symposium / XXI National Congress on Solid State Dosimetry. Hermosillo, Son., Mexico, 28-2/Oct/2009.

[231] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., and Mendez-Villafane

R. Comparacion del espectro de neutrones de un detector de 3He obtenido con dos codigos de deconvolucion. In 11 Seminario de Investigacion UAA. Aguascalientes, Ags.,
Mexico, 18-21/May/2010.
[232] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. PerFillol A., Mendez-Villafane
formance of artificial neutral networks and genetical evolved artificial neutral networks
unfolding techniques. In VI International Symposium on Radiation Physics. Zacatecas,
Zac., Mexico, 7-10/Mar/2010.
[233] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. ComFillol A., Mendez-Villafane
paring the neutron spectra unfolding features using two codes. In VI International
Symposium on Radiation Physics. Zacatecas, Zac., Mexico, 7-10/Mar/2010.
[234] Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. A new computer
tool based on ANN technology, applied in neutron spectrometry and dosimetry research
areas. Proceedings of the Electronics, Robotics and Automotive Mechanics Conference
(CERMA08), IEEE Computer Society, pages 189194, 2008.
[235] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A computational tool design for evolutionary artificial neural networks in neutron spectrometry and dosimetry. Proceedings of the Electronics, Robotics and Automotive Mechanics
Conference (CERMA09), IEEE Computer Society, pages 113118, 2009.
[236] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A neutron spectrometry and dosimetry computer tool based on ann. In 12th Congress of
the International Radiation Protection Association (IRPA12), pages xxyy, International
Atomic Energy Agency, Vienna, Oct 2010. IAEA Procceding Series STI/PUB/1460 1-9.

BIBLIOGRAFIA

224

[237] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., and Mendez-Villafane

R. Espectrometra y dosimetra neutronicas con redes neuronales artificiales para un
detector de 3He. In 11 Seminario de Investigacion UAA. Aguascalientes, Ags., Mexico,
18-21/May/2010. Trabajo distinguido como la mejor ponencia.
[238] Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Los Arcos-Merino
R. Calculo

J.M., and Mendez-Villafane

del error producido por las lecturas de conteo en
la espectrometra neutronica por medio de esferas Bonner y la distribucion de poisson.
In 11 Seminario de Investigacion UAA. Aguascalientes, Ags., Mexico, 18-24/May/2010.
[239] IAEA. Compendium of neutron spectra and detector responses for radiation protection
purposes. Technical Report 403, 2001.
[240] ICRP. Conversion coefficients for use in radiological protection against external radiation.icrp74. International Commission on Radiological Protection, 26(3-4):199200,
1996.
[241] BiocompSystems. www.biocompsystems.com. May 2011.
[242] NI. Manual de referencia de funciones e IV. National Instruments.
[243] NI. Manual de referencia de programacion en g. National Instruments.
[244] NI. Que son las mediciones y la instrumentacion virtual. National Instruments.

[245] Correa J.C. Estimacion por intervalo del parametro

de la distribucion de poisson por
una sola observacion. Revista Colombiana de Estadstica, 30(1):6975, 2007.
[246] Walecka J.D. Fundamentals of statistical mechanics. World Scientific, 2000.
[247] Soong T.T. Fundamentals of probability and statistics for engineers. John Wiley & Sons,
Inc., 2004.
[248] Singh Y.K. Fundamentl of research methodology and statistics. New Age International
Publishers, 2006.

Jos Manuel
Ortiz Rodrguez
Docente-Investigador UAIE-UAZ

Universidad Autnoma de Zacatecas,

Unidad Acadmica de Ingeniera Elctrica,
Av. Ramn Lpez Velarde # 801, Col.
Centro, Zacatecas, Zacatecas, Mxico
98000
T +52 (492) 923-9407 Ext. 1515 y/o 1517
B morvymm@yahoo.com.mx

Curriculum Vitae

Idiomas
Ingls

Alto

540 puntos TOEFL

Experiencia Laboral
2004Presente

Docente-investigador, Universidad Autnoma de Zacatecas (UAZ), Zacatecas.

Docente-Investigador de la Unidad Acadmica de Ingeniera Elctrica (UAIE) de la UAZ.

Sistemas CAD-CAM. Ene-Jul/2010,2011;
Dibujo computarizado Ago-Dic/2008, Ene-Jul/2009,2010,2011;
Cursos de nivelacin en lgebra y clculo para alumnos de nuevo ingreso. Jul-Ago/2010;
Dispositivos Lgicos Programables. Ene-Jul/2009;
Electrnica digital y laboratorio. 2004 - 2010;
Cursos de nivelacin en computacin para alumnos de nuevo ingreso. Jun-Jul/2005, 2006;
Coordinador del Centro de Cmputo del rea de Ingeniera en Comunicaciones y Electrnica.
Feb/2004Jun/2008;

Nov. 2002
Abril 2003

Capturista de datos, Instituto Federal Electoral (IFE), Zacatecas.

Ene. 2000
Mayo 2001

Ingeniero de Mantenimiento, Sistemas y electrnica, Zacatecas.

Mar. 1999
Nov. 2002

Auxiliar, Centro de Cmputo del rea de Ingeniera (CECAI) - (UAZ), Zacatecas.

Captura de la informacin generada en el proceso electoral federal 2003.

Planeacin, diseo y desarrollo de proyectos de redes de cmputo; instalacin y configuracin

de software y hardware; mantenimiento preventivo y correctivo de equipo de cmputo.
Participacin directa en la planeacin y diseo de la reestructuracin de la red de cmputo e
instalaciones del CECAI e impartiendo los cursos:

Instalacin, configuracin y administracin de redes locales con Windows 2000. Jul/2003;

Instalacin, configuracin y administracin de redes locales con Windows NT. Jul/2002;
Mantenimiento preventivo y correctivo de PCs e instalacin y configuracin de software y
hardware. Dic/2000Jul/2001;
Windows, Word, Excel, Power Point e Internet. MarMay/2000, MarMay/2001 y Jun/2001;

1/12

Ago. 1994
Ene. 1999

Cabo de transmisiones, Ejrcito Mexicano, Mxico.

Actividades castrenses en el Ejrcito Mexicano, desempeando los siguientes cargos:

Dragn del 22 Regimiento de Caballera Motorizado (22 RCM), en fuerzas bsicas, con el
grado de soldado de caballera, granadero de fusil y proveedor;
Miembro del grupo especial pelotn de armas de apoyo con el grado de soldado proveedor;
Soldado del servicio de transmisiones militares, proporcionando asesora y apoyo logstico;
Miembro del grupo elite de fuerzas especiales Pelotn de Comunicaciones Especiales, realizando actividades de inteligencia en el rea de comunicaciones militares, con el grado de
cabo de transmisiones;

Capacidad tcnica
1

Sistemas de clculo cientfico: Matlab, Mathcad, Mathematica.

Configuracin y administracin del

Sistema Operativo Linux (Debian y
Suse).

El diseo de programas con los

lenguajes de programacin: LabVIEW, Visual Basic, Delphi, C++.

Diseo y desarrollo de sistemas electrnicos con: Microcont., FPGA y

VHDL.

Varios sistemas operativos de servidores como: Windows NT 4, Windows 2000, Windows 2003, Linux
SUSE, Linux Debian.
Diseo, instalacin, configuracin,
administracin y mantenimiento de:
redes de cmputo, servidores y
equipo de cmputo.

Amplio conocimiento y dominio del

paquete computacional Office.

Diseo en: Corel Draw, Autocad,

Front Page.

Manejo de sistemas de bases de datos

con Access y SQL.

La teora de diseo robusto aplicando

metodologa Taguchi.

Varios Sistemas Operativos: desde

MS-DOS, Windows 95 hasta Win7.

La teora de clculo paralelo en redes

de computadoras no dedicadas usando PVM.

20072011

Paquetes de diseo electrnico como:

Orcad, Circuit Maker, Multisim, Xilinx, Max Plus II.
La teora de redes neuronales artificiales y algoritmos genticos.

Formacin acadmica

Estudiante de doctorado en el programa: Tcnicas Avanzadas de Anlisis,

Simulacin y Control de Sistemas, Universidad de Crdoba, Crdoba, Espaa,
Doctorado.

Dic. 2005

Maestra en Ciencias Nucleares, con opcin terminal en instrumentacin nuclear, antes Centro Regional de Estudios Nucleares, ahora UAEN (UAZ), Zacatecas,
Maestra, C.P. 5311390 .

Jun. 2003

Ingeniero en Comunicaciones y Electrnica, antes Facultad de Ingeniera, ahora

UAIE (UAZ), Zacatecas, Licenciatura, C.P. 3715361 .

Oct. 2009
Mar. 2010

Estancia de investigacin pre-doctoral en el CIEMAT por medio de una beca

concedida por la AECID, para realizar los trabajos de la tesis doctoral efectuada
en la Universidad de Crdoba en Espaa, Lab. Patr. Neutr. Lab. Metrol. Rad.
Ioniz. Ctro. Inv. Medio Amb. Energ. y Tecnol. (CIEMAT) , Madrid, Espaa, Estancia
pre-doctoral.

2/12

Jul. 2009

Taller de Dosimetra Termoluminiscente, Escuela de verano UAZ-SPAUAZ 2009,

Zacatecas, Diplomados.

Oct. 2008
Mar. 2009

Diplomado en diseo de sistemas digitales (VHDL), Universidad Autnoma de

Zacatecas Universidad de Quertaro, Zacatecas, Diplomados.

Oct. 2007
Nov. 2007

Prcticas profesionales en el Hospital General de Mxico, en la caracterizacin

de materiales para dosimetra TLD en aplicaciones mdicas., Hospital General
de Mxico, Mxico D.F., Prcticas profesionales de maestra.

Oct. 2005 Dic.

2005

Diplomado en diseo de sistemas digitales (VHDL), Universidad Politcnica de

Aguascalientes (UPA), Aguascalientes, Diplomados.

Sep. 2005

Diplomado en dosimetra termoluminiscente aplicada a la medicina, Universiadad Autnoma de Mxico - Universidad Autnoma de Zacatecas, Zacatecas, Diplomados.

Jun. 2004
Ago. 2004

Diplomado formacin de investigadores, Centro de Biologa Experimental (UAZ),

Zacatecas, Diplomados.

Mayo 2004

Prcticas profesionales en el reactor nuclear del Instituto Nacional de Investigaciones Nucleares, Instituto Nacional de Investigaciones Nucleares (ININ), Toluca,
estado de Mxico, Prcticas profesionales maestra.

Jul. 2002

Curso de Instalacin, configuracin y administracin de servidores Linux , Centro de cmputo del rea de ingeniera CECAI, (UAZ), Zacatecas, Cursos.

2000 2002

Curso bsico e intermedio de ingls , KELLER, Zacatecas, Curso Ingls.

Oct. 2000 Dic.

2000

Curso de programacin en lenguaje Delphi, Centro de cmputo del rea de ingeniera CECAI, (UAZ), Zacatecas, Cursos.

Mar. 2000
Abr. 2000

Curso de diseo grfico con Corel Draw, Centro de cmputo del rea de ingeniera
CECAI, (UAZ), Zacatecas, Cursos.

Ene. 1999
Ene. 1999
Sep. 1998

Curso de Windows, Office y Multimedia, CETEC, Hermosillo, Sonora, Cursos.

Curso bsico de ingls, CETEC, Hermosillo, Sonora, Cursos.
Adiestramiento en Inteligencia en Comunicaciones Militares, EJRCITO
MEXICANO, Mexicali, B.C.N., Cursos.

Tesis de Doctorado
Ttulo:
Asesores:

Ttulo:
Asesores:

Ttulo:

Asesor:

Inteligencia Artificial: aplicaciones en la espectrometra y dosimetra neutrnicas

Vega-Carrillo H.R. y Benavides-Bentez J.I.

Tesis de Maestra

Diseo robusto de redes neuronales artificiales aplicadas en la espectrometra de

neutrones
Vega-Carrillo H.R. y Hernndez-Dvila V.M.

Tesis de Licenciatura

Diseo de sistemas de adquisicin de datos en la instrumentacin virtual, empleando

dispositivos estndar de la industria electrnica y el lenguaje grfico de programacin
Labview
Cervantes-Viramontes J.M.

3/12

Participacin en Congresos
Internacionales
Mar. 2010

International Symposium on Radiation Physics, Universidad Autnoma de Zacatecas (UAZ), Zacatecas, Zac., Mxico.

Nov. 2009

American Nuclear Society, American Nuclear Society: 2009 Winter Meeting,

Whashintong, USA.

Oct. 2009

XI International Symposium / XXI National Congress on Solid State Dosimetry, Universidad de Sonora, Hermosillo, Son., Mxico.

Sep. 2009

Electronics, Robotics and Automotive Mechanics Conference, CERMA09, Cuernavaca, Mor., Mxico.

Oct. 2008

XII congreso Internacional de la Sociedad Internacional de Proteccin Radiolgica, IRPA08, Buenos Aires, Argentina.

Sep.Oct. 2008

Electronics, Robotics and Automotive Mechanics Conference, CERMA08, Cuernavaca, Mor., Mxico.

Sep. 2007

X International Symposium / XX National Congress on Solid State Dosimetry,

Universidad Autnoma de Puebla, Puebla, Pue., Mxico.

Sep. 2006

Electronics, Robotics and Automotive Mechanics Conference, CERMA06, Cuernavaca, Mor., Mxico.

Sep. 2006

XX Congreso Anual de la Sociedad Nuclear Mexicana y Congreso Internacional

dela Sociedad Internacional de Proteccin Radiolgica (IRPA), SNM-IRPA, Acapulco, Gro., Mxico.

Ago.Sep. 2006

IX International Symposium / XIX National Congress on Solid State Dosimetry, Instituto Politcnico Nacional (IPN), Mxico D.F., Mxico.

Sep. 2005

VIII International Symposium / XVIII National Congress on Solid State

Dosimetry, Universidad Autnoma de Zacatecas (UAZ), Zacatecas, Zac., Mxico.

Nacionales
Jul. 2009

XX Congreso Anual de la Sociedad Nuclear Mexicana, SNM, Puerto Vallarta,

Jal., Mxico.

Regionales
Ago. 2010

11o seminario de investigacin, Universidad Autnoma de Aguascalientes (UAA),

Aguascalientes, Ags., Mxico.

Ago. 2009

10o seminario de investigacin, Universidad Autnoma de Aguascalientes (UAA),

Aguascalientes, Ags., Mxico.

Mayo 2008

9o seminario de investigacin, Universidad Autnoma de Aguascalientes (UAA),

Aguascalientes, Ags., Mxico.

Locales
Mayo. 2008

XII Jornadas de Investigacin, Universidad Autnoma de Zacatecas (UAZ), Zacatecas, Zac., Mxico.

Ago. 2007

XI Jornadas de Investigacin, Universidad Autnoma de Zacatecas (UAZ), Zacatecas, Zac., Mxico.

4/12

Abr. 2006

Encuentro de Investigacin de Ingeniera Elctrica (ENINVIE06), Universidad

Autnoma de Zacatecas (UAZ), Zacatecas, Zac., Mxico.

Ago. 2004

Jornadas de la investigacin de la Unidad Acadmica de Ciencias Qumicas

(UAZ), Universidad Autnoma de Zacatecas (UAZ), Zacatecas, Zac., Mxico.

Premios y Distinciones
1

Se obtuvo una beca ayuda tesis por parte de PROMEP comprendida entre el periodo mayo
de 2010 a abril de 2011.

El trabajo Espectrometra y dosimetra neutrnica con redes neuronales artificiales para un

detector de 3He obtenido con dos cdigos de deconvolucin, presentado en el XI Seminario de
Investigacin de la Universidad Autnoma de Aguascalientes en 2010, fue seleccionado como
la Mejor Ponencia en su mesa temtica.

Se obtuvo una beca MAEC-AECID de octubre de 2009 a marzo de 2010, para realizar una
estancia de investigacin pre-doctoral en el centro de investigacin CIEMAT de Espaa, en
colaboracin con la UCO, UAZ y UPM con el objeto de realizar los trabajos correspondientes
al tema de tesis doctoral a ser desarrollados en este programa de doctorado.

En la Unidad Acadmica de Estudios Nucleares (UAEN) de la Universidad Autnoma de

Zacatecas (UAZ), la M.C. Ma. Del Rosario Martnez Blanco, realiz la tesis de Maestra
denominada NSDann: una herramienta de cmputo para la espectrometra y dosimetra de
neutrones basada en RNA, Asesorada por el Dr. Hctor Rene Vega Carrillo y el que escribe.
El examen profesional se realiz en Julio de 2009, obtenindose la distincin de Mencin
Honorfica, siendo sta la primera y nica otorgada en la historia de la UAEN de la UAZ.

El trabajo presentado en el congreso de la Sociedad Nuclear Mexicana en 2009, denominado

Una nueva herramienta de cmputo para la espectrometra y dosimetra de neutrones, fue
seleccionado el como mejor trabajo de la mesa y se public en la revista Mxico Nuclear.

El trabajo presentado en el congreso regional:10 seminario de Investigacin de la Universidad

Autnoma de Aguascalientes denominado Una nueva herramienta de cmputo para la espectrometra y dosimetra de neutrones, fue seleccionado como el mejor trabajo de la mesa de
Ingenieras II.

El trabajo presentado en el congreso regional: 10 seminario de Investigacin de la Universidad

Autnoma de Aguascalientes denominado Una nueva propuesta para dosimetra neutrnica por
medio de mtodos iterativos, fue seleccionado como el mejor trabajo de la mesa de Ingenieras
I.

Se obtuvo apoyo por parte de la OMS, para asistir al XII congreso Internacional de la Sociedad
Internacional de Proteccin Radiolgica (IRPA08), Buenos Aires, Argentina, Oct-2008.

Mejor trabajo en presentacin oral en la XVIII Conferencia Internacional y VIII Congreso

Nacional sobre dosimetra de estado slido, Zacatecas, Zac., 2005, con el tema Reconstruccin
de espectros de neutrones a partir del sistema espectromtrico de esferas de Bonner.

Mejor trabajo en presentacin cartel en la XVIII Conferencia Internacional y VIII Congreso

Nacional sobre dosimetra de estado slido, Zacatecas, Zac., 2005, con el tema Aplicacin de
dosmetros TLD en salas de Rayos-X.

5/12

Membresas
AJENM

Asociacin de Jvenes por la Energa Nuclear en Mxico

NA-YGN

Nuclear American Young Generation in Nuclear

SMID

Sociedad Mexicana de Irradiacin y Dosimetra

SNM

Sociedad Nuclear Mexicana

Tesis
Dirigidas
1

Martnez-Blanco M.R.Espectrometra de neutrones y clculo simultneo de dosis equivalentes

empleando la metodologa de diseo robusto de redes neuronales. Tesis de licenciatura. Unidad
Acadmica de Ingeniera Elctrica, UAZ. 2006.

Martnez-Blanco M.R. NSDann: una herramienta de cmputo para la espectrometra y

dosimetra de neutrones basada en RNA. Tesis de maestra. Unidad Acadmica de Ciencias
Nucleares, UAZ. 2009.

Evaluadas
1

Veyna-Lamas M. Proceso de produccin de software en una empresa. Tesis de licenciatura.

Unidad Acadmica de Ingeniera Elctrica, UAZ. 2007.

Torres-Ramrez J.G. Determinacin de la actividad de una fuente por el mtodo relativo. Tesis
de maestra. Unidad Acadmica de Ciencias Nucleares, UAZ. 2008.

Hernndez-Ortiz M. Kerma en aire y dosis absorvida en el torso humano durante un estudio

mastogrfico. Tesis de maestra. Unidad Acadmica de Ciencias Nucleares, UAZ. 2009.

Durn-Muoz H.A. textitDosis glandular media y kerma en aire en mamografa. Tesis de

maestra. Unidad Acadmica de Ciencias Nucleares, UAZ. 2009.

Pinedo-Sols A. Dosis en radiografa de trax. Tesis de maestra. Unidad Acadmica de Ciencias

Nucleares, UAZ. 2009.

Revisadas
1

Arcos-Pichardo A. Dosis en rganos internos de una mujer embarazada debida al 131I. Tesis
de maestra. Unidad Acadmica de Ciencias Nucleares, UAZ. 2008.

Hernndez-Villasana R. Caracterizacin de una fuente de neutrones de 239PuBe. Tesis de

maestra. Unidad Acadmica de Ciencias Nucleares, UAZ. 2008.

Chacn-Ruz A. Espectrmetro de partculas [U+F061] con un fotodiodo tipo PIN. Tesis de

maestra. Unidad Acadmica de Ciencias Nucleares, UAZ. 2009.

Patentes y derechos de autor

[1]

Ortiz-Rodrguez J.M. Diseo robusto de redes neuronales artificiales aplicadas en la espectrometra de neutrones. 03-2006-031312565200-01, 31/Mar/2006.

[2]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M. DRRNA Ver. 1.3. 03-2006-031312574800-01,

31/Mar/2006.

[3]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M.

061611050300-01, 30/Jun/2008.

Reconstructor Ver. 1.0.

03-2008-

6/12

[4]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M. DRRNA Ver. 2.14. 03-2008-061611072500-01,

30/Jun/2008.

[5]

Martnez-Blanco M.R. Espectrometra de neutrones y cluclo simultneo de dosis equivalentes

empleando la metodologa de diseo robusto de redes neuronales. 03-2008-061611102500-01,
30/Jun/2008.

[6]

Martnez-Blanco M.R. NSDann: una herramienta de cmputo para la espectrometra y

dosimetra de neutrones basada en rna. 03-2010-011911330600-01, 4/Feb/2010.

[7]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M. VALGRAF Ver. 1.0. 03-2010-01191133550001, 4/Feb/2010.

[8]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M. NSDANN Ver. 0.1.1. 03-2010-01191135000001, 4/Feb/2010.

[9]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M. VMCNP Ver. 0.1.1. 03-2010-01191136400001, 4/Feb/2010.

[10]

Martnez-Blanco M.R. and Ortiz-Rodrguez J.M. TBGANN Ver. 0.1.1. 03-2010-01191137530001, 4/Feb/2010.

[11]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDUAZ Ver. 0.1.0. 032010-011911390300-01, 4/Feb/2010.

[1]

Publicaciones en libros

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Accepted to be published

in Evolutionary artificial neural networks in neutron spectrometry. Chapter: Artificial neural
networks. Ed. INTECH. 2010. ISBN 978-953-307-243-2.

Publicaciones en revistas

[1]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Robust design of artificial
neural networks applying the Taguchi methodology and DoE. Proceedings of the Electronics,
Robotics and Automotive Mechanics Conference (CERMA06), IEEE Computer Society, 2006.

[2]

Vega-Carrillo H.R., Hernndez-Dvila V.M., Manzanares-Acua E., Mercado-Snchez G.A.,

Iiguez de la Torre M.P., Barquero R., Preciado-Flores S., Mndez-Villafae R., Arteaga-Arteaga
T., and Ortiz-Rodrguez J.M. Neutron spectrometry using artificial neural networks. Radiation
Measurements, 41:425431, 2006.

[3]

Vega-Carrillo H.R., Manzanares-Acua E., Ortiz-Rodrguez J.M., and Arteaga-Arteaga T. Neutron spectra re-binning and dose calculation using Monte Carlo methods. Revista Mexicana de
Fsica, 53(3):17, 2007.

[4]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. A new computer tool
based on ANN technology, applied in neutron spectrometry and dosimetry research areas. Proceedings of the Electronics, Robotics and Automotive Mechanics Conference (CERMA08), IEEE
Computer Society, pages 189194, 2008.

[5]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Artificial neural
networks modeling evolved genetically, a new approach applied in neutron spectrometry and
dosimetry research areas. Proceedings of the Electronics, Robotics and Automotive Mechanics
Conference (CERMA08), IEEE Computer Society, pages 387392, 2008.

[6]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann, a LabVIEW tool
for neutron spectrometry and dosimetry based on the RDANN methodology. Proceedings of
the Electronics, Robotics and Automotive Mechanics Conference (CERMA09), IEEE Computer
Society, 00:131136, 2009.

[7]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A computational tool design for evolutionary artificial neural networks in neutron spectrometry and dosimetry.

7/12

Proceedings of the Electronics, Robotics and Automotive Mechanics Conference (CERMA09),

IEEE Computer Society, pages 113118, 2009.
[8]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann: Espectrometra

y dosimetra de neutrones. Mxico Nuclear, 9(2):1929, 2009.

[9]

Vega-Carrillo H.R., Martnez-Blanco M.R., Hernndez-Dvila V.M., and Ortiz-Rodrguez J.M.

Spectra and dose with ANN of 252Cf, 241AmBe and 239PuBe. Journal of Radianalitycal and
Nuclear Chemistry, 281(3):615618, 2009.

[10]

Vega-Carrillo H.R., Martnez-Blanco M.R., Hernndez-Dvila V.M., and Ortiz-Rodrguez J.M.

Spectra and dose with ANN of 252Cf, 241Am-Be, and 239Pu-Be. Journal of Radioanalytical
and Nuclear Chemistry, 281(3):615618, 2009.

[11]

Mndez-Villafae R., Guerrero-Araque J.E., Gallego E., Lorente-Fillol A., Tore C., OrtizRodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. 241Am-Be source term simulation with Monte Carlo methods. accepted to be published in: Revista Mexicana de Fsica,
2010.

[12]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. Comparing the neutron
spectra unfolding features using two codes. accepted to be published in: Revista Mexicana de
Fsica, 2010.

[13]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. Performance of artificial
neutral networks and genetical evolved artificial neutral networks unfolding techniques. accepted
to be published in: Revista Mexicana de Fsica, 2010.

[14]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. User-friendly interface
for run MCNP code for radiation physics. accepted to be published in: Revista Mexicana de
Fsica, 2010.

[15]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. Neutron spectrometry
using artificial neutral networks for a Bonner sphere spectrometer with a He-3 detector. accepted
to be published in: Revista Mexicana de Fsica, 2010.

[16]

Vega-Carrillo H.R., Ortiz-Rodrguez J.M., Hernndez-Dvila V.M., Martnez-Blanco M.R.,

Hernndez-Almarz B., Ortiz-Hernndez A.A., and Mercado-Snchez G.A. Different spectra
with the same neutron source. Revista Mexicana de Fsica, 51(6):3539, 2010.

[17]

Vega-Carrillo H.R., Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Hernndez-Dvila V.M.

ANN in spectrometry and neutron dosimetry. American Institute of Physics Proccedings,
1310:1217, 2010.

Publicaciones en congresos
[1]

Hernndez-Dvila V.M., Vega-Carrillo H.R., Manzanares-Acua E., Mercado-Snchez G.A.,

Iiguez de la Torre M.P., Barquero R., Arteaga-Arteaga T., and Ortiz-Rodrguez J.M. Neutron
spectrometry using artificial neural networks. In VIII Jornadas de Investigacin UAZ. Zacatecas,
Zac., Mxico, 10-12/Nov/2004.

[2]

Vega-Carrillo H.R., Hernndez-Dvila V.M., Manzanares-Acua E., Mercado-Snchez G.A.,

Arteaga-Arteaga T., and Ortiz-Rodrguez J.M. Neutron spectra unfolding with artificial neural
networks. In Encuentro de Investigacin en Ingeniera Elctrica (ENINVIE-UAZ2005). Zacatecas, Zac., Mxico, 17-18/Mar/2005.

[3]

Vega-Carrillo H.R., Hernndez-Dvila V.M., Manzanares-Acua E., Mercado-Snchez G.A.,

Ortiz-Rodrguez J.M., Iiguez de la Torre M.P., Barquero R., and Arteaga-Arteaga T. Neutron spectrometry with artificial neural networks. In XVI Congreso Anual de la SNM y XXIII
Reunin Anual de la SMSR. Oaxaca, Oax., Mxico, 10-13/Jul/2005.
8/12

[4]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Arteaga-Arteaga T., Vega-Carrillo H.R.,

Hernndez-Dvila V.M., and Manzanares-Acua E. Reconstruccin de espectros de neutrones
usando redes neuronales a partir del sistema espectromtrico de esferas de Bonner. In VIII Conferencia Internacional / XVIII Congreso Nacional sobre Dosimetra de Estado Slido. Zacatecas,
Zac., Mxico, 7-9/Sep/2005.

[5]

Letechipia de Len C., Lara-Zaldo L.E.C., Perales-Muoz W., Preciado-Flores S., and OrtizRodrguez J.M. Aplicacin de dosmetros TLD en salas de rayos X. In VIII Conferencia Internacional / XVIII Congreso Nacional sobre Dosimetra de Estado Slido. Zacatecas, Zac., Mxico,
7-9/Sep/2005.

[6]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Arteaga-Arteaga T., Vega-Carrillo H.R.,

Hernndez-Dvila V.M., and Manzanares-Acua E. Diseo robusto de redes neuronales artificiales multicapa de propagacin hacia delante, con el algoritmo de propagacin inversa, en
la reconstruccin de espectros de neutrones. In IX Jornadas de Investigacin UAZ. Zacatecas,
Zac., Mxico, 9-11/Nov/2005.

[7]

Arteaga-Arteaga T., Ortiz-Rodrguez J.M., Vega-Carrillo H.R., Hernndez-Dvila V.M.,

Manzanares-Acua E., and Mercado-Snchez G.A. Diseo de una red neuronal artificial, con la
topologa orientada a la deconvolucin del espectro de neutrones. In IX Jornadas de Investigacin
UAZ. Zacatecas, Zac., Mxico, 9-11/Nov/2005.

[8]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Diseo robusto de redes
neuronales artificiales. In Encuentro de Investigacin en Ingeniera Elctrica (ENINVIE-UAZ
2006). Zacatecas, Zac., Mxico, 5-7/Abr/2006.

[9]

Ortiz-Rodrguez J.M. and Martnez-Blanco M.R. Diseo de neuro-hadware. In Encuentro

de Investigacin en Ingeniera Elctrica (ENINVIE-UAZ2006). Zacatecas, Zac., Mxico, 57/Abr/2006.

[10]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Espectrometra de neutrones y clculo de dosis equivalentes, aplicando la metodologa de diseo robusto de redes neuronales artificiales. In Encuentro de Investigacin en Ingeniera Elctrica (ENINVIE-UAZ2006).
Zacatecas, Zac., Mxico, 5-7/Abr/2006.

[11]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Problemas crticos en

el diseo y construccin de redes neuronales, en la espectrometra de neutrones, utilizando la
metodologa de diseo robusto de redes neuronales artificiales. In Primer Congreso Americano
del IRPA / XXIV Reunin Anual de la SMSR / XVII Congreso Anual de la SNM. Acapulco,
Gro., Mxico, 3-8/Sep/2006.

[12]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Neutron spectrometry

and dosimetry by means of Bonner spheres system and artificial neural networks applying robust
design of artificial neural networks. In Primer Congreso Americano del IRPA / XXIV Reunin
Anual de la SMSR / XVII Congreso Anual de la SNM. Acapulco, Gro., Mxico, 3-8/Sep/2006.

[13]

Arteaga-Arteaga T., Ortiz-Rodrguez J.M., Vega-Carrillo H.R., Hernndez-Dvila V.M.,

Manzanares-Acua E., and Mercado-Snchez G.A. Diseo de una red neuronal artificial, con la
topologa orientada a la reconstruccin del espectro de neutrones. In Primer Congreso Americano del IRPA / XXIV Reunin Anual de la SMSR / XVII Congreso Anual de la SNM. Acapulco,
Gro., Mxico, 3-8/Sep/2006.

[14]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. RDANN a new methodology to solve the neutron spectra unfolding problem. In IX International Symposium / XIX
Nacional Meeting on Solid State Dosimetry. Mxico, D.F, 29-1/Sep/2006.

[15]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. A proposal to order the
neutron data set in neutron spectrometry using the RDANN methodology. In IX International
Symposium / XIX Nacional Meeting on Solid State Dosimetry, pages 209214. Mxico, D.F,
29-1/Sep/2006.

[16]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Neutron spectrometry

and dosimetry by means of bonner spheres system and artificial neural networks applying robust
9/12

design of artificial neural networks. In IX International Symposium / XIX Nacional Meeting on

Solid State Dosimetry. Mxico, D.F, 29-1/Sep/2006.
[17]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Robust design of artificial
neural networks applying the Taguchi methodology and DoE. In Congreso de Electrnica,
Robtica y Mecnica Automotriz CERMA 2006. Cuernavaca, Mor., Mxico, 26-29/Sep/2006.

[18]

Ortiz-Rodrguez J.M., Vega-Carrillo H.R., and Martnez-Blanco M.R. ANN knowledge extraction
at Matlab environment, for neutron spectrometry and simultaneous dosimetry. In XI Jornadas
de Investigacin UAZ. Zacatecas, Zac., Mxico, 22-24/Ago/2007.

[19]

Ortiz-Rodrguez J.M. and Martnez-Blanco M.R. ANN optimization by means of RDANN

and RSM methodologies. In XI Jornadas de Investigacin UAZ. Zacatecas, Zac., Mxico, 2224/Ago/2007.

[20]

Ortiz-Rodrguez J.M. and Martnez-Blanco M.R. Genetic artificial neural networks (GANN)
modeling. In XI Jornadas de Investigacin UAZ. Zacatecas, Zac., Mxico, 22-24/Ago/2007.

[21]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Development of a new

software tool, based on ANN technology, in neutron spectrometry and dosimetry research. In
X International Symposium / XX National Congress on Solid State Dosimetry. Puebla, Pue.,
Mxico, 24-27/Sep/2007.

[22]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Diseo de redes neuronales
artificiales por medio de algortmos genticos, aplicadas en la espectrometra y dosimetra de neutrones. In 9o seminario de investigacin UAA. Aguascalientes, Ags., Mxico, 20-22/May/2008.

[23]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Espectrometra y

dosimetra de neutrones por medio de redes neuronales evolutivas. In XIX Congreso Anual
de la SNM. Mrida, Yuc., Mxico, 6-9/Jul/2008.

[24]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Diseo de redes neuronales
artificiales por medio de algoritmos genticos, aplicadas en la espectrometra y dosimetra de
neutrones. In XII Jornadas de Investigacin UAZ. Zacatecas, Zac., Mxico, 13-15/Ago/2008.

[25]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Diseo de una interfase
grfica de usuario, para la espectrometra y dosimetra de neutrones, basado en redes neuronales
artificiales. In XII Jornadas de Investigacin UAZ. Zacatecas, Zac., Mxico, 13-15/Ago/2008.

[26]

Ortiz-Rodrguez J.M. and Martnez-Blanco M.R. Solucin de problemas en ingeniera por medio
de algoritmos genticos, descripcin de un ejemplo sencillo de optimizacin. In XII Jornadas de
Investigacin UAZ. Zacatecas, Zac., Mxico, 13-15/Ago/2008.

[27]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Artificial neural
networks modeling evolved genetically, a new approach applied in neutron spectrometry and
dosimetry research areas. In Congreso de Electrnica, Robtica y Mecnica Automotriz CERMA
2008. Cuernavaca, Mor., Mxico, 30-3/Oct/2008.

[28]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. A new computer tool
based on ANN technology, applied in neutron spectrometry and dosimetry research areas. In
Congreso de Electrnica, Robtica y Mecnica Automotriz CERMA 2008. Cuernavaca, Mor.,
Mxico, 30-3/Oct/2008.

[29]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Neutron spectrometry and dosimetry based on a new approach called genetic artificial neural networks. In 12th
International Congress of the International Radiation Protection Association (IRPA). Buenos
Aires, Argentina, 19-24/Oct/2008.

[30]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A neutron
spectrometry and dosimetry computer tool based on ANN. In 12th International Congress
of the International Radiation Protection Association (IRPA). Buenos Aires, Argentina, 1924/Oct/2008.

10/12

[31]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Diseo de
una herramienta de cmputo para la espectrometra y dosimetra de neutrones por medio de
redes neuronales evolutivas. In XX Congreso Anual de la SNM. Puerto Vallarta, Jal., Mxico,
5-8/Jul/2009.

[32]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Una nueva herramienta de
cmputo para la espectrometra y dosimetra de neutrones. In XX Congreso Anual de la SNM.
Puerto Vallarta, Jal., Mxico, 5-8/Jul/2009. Trabajo distinguido como la mejor ponencia.

[33]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Diseo de una herramienta
de cmputo para la espectrometra y dosimetra de neutrones con redes neuronales evolutivas.
In 10o seminario de investigacin UAA. Aguascalientes, Ags., Mxico, 26-28/Ago/2009.

[34]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. Una nueva propuesta para
dosimetra neutrnica por medio de mtodos iterativos. In 10o seminario de investigacin UAA.
Aguascalientes, Ags., Mxico, 26-28/Ago/2009. Trabajo distinguido como la mejor ponencia.

[35]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann espectrometra

y dosimetra de neutrones por medio de redes neuronales artificiales. In 10o seminario de investigacin UAA. Aguascalientes, Ags., Mxico, 26-28/Ago/2009. Trabajo distinguido como la
mejor ponencia.

[36]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. A computational tool design for evolutionary artificial neural networks in neutron spectrometry and dosimetry.
In Congreso de Electrnica, Robtica y Mecnica Automotriz CERMA 2009. Cuernavaca, Mor.,
Mxico, 22-25/Sep/2009.

[37]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. NSDann, a LabVIEW

tool for neutron spectrometry and dosimetry based on the RDANN methodology. In Congreso
de Electrnica, Robtica y Mecnica Automotriz CERMA 2009. Cuernavaca, Mor., Mxico,
22-25/Sep/2009.

[38]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Espectrometra
de las fuentes de neutrones de 239PuBe, 252Cf y 241AmBe por medio de NSDUAZ. In XI
International Symposium / XXI National Congress on Solid State Dosimetry. Hermosillo, Son.,
Mxico, 28-2/Oct/2009.

[39]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Una comparacin en la reconstruccin de espectros de neutrones utilizando tcnicas iterativas clsicas y las
basadas en inteligencia artificial. In XI International Symposium / XXI National Congress on
Solid State Dosimetry. Hermosillo, Son., Mxico, 28-2/Oct/2009.

[40]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Redes neuronales evolutivas: una nueva alternativa para la espectrometra de neutrones. In XI International
Symposium / XXI National Congress on Solid State Dosimetry. Hermosillo, Son., Mxico, 282/Oct/2009.

[41]

Martnez-Blanco M.R., Ortiz-Rodrguez J.M., and Vega-Carrillo H.R. Espectrometra y

dosimetra de neutrones en aceleradores de partculas de uso mdico, utilizando NSDANN.
In XI International Symposium / XXI National Congress on Solid State Dosimetry. Hermosillo,
Son., Mxico, 28-2/Oct/2009.

[42]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Gallego E., and Vega-Carrillo H.R. Evolutionary
artificial neural networks in neutron dosimetry. In American Nuclear Society: 2009 Winter
Meeting. Washington, D.C., EEUU, 15-19/Nov/2009.

[43]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. Comparing the neutron
spectra unfolding features using two codes. In VI International Symposium on Radiation Physics.
Zacatecas, Zac., Mxico, 7-10/Mar/2010.

[44]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. Performance of artificial
11/12

neutral networks and genetical evolved artificial neutral networks unfolding techniques. In VI
International Symposium on Radiation Physics. Zacatecas, Zac., Mxico, 7-10/Mar/2010.
[45]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. User-friendly interface
for run MCNP code for radiation physics. In VI International Symposium on Radiation Physics.
Zacatecas, Zac., Mxico, 7-10/Mar/2010.

[46]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Gallego E., Lorente-Fillol A.,
Mndez-Villafae R., Los Arcos-Merino J.M., and Guerrero-Araque J.E. Neutron spectrometry
using artificial neutral networks for a Bonner sphere spectrometer with a He-3 detector. In VI
International Symposium on Radiation Physics. Zacatecas, Zac., Mxico, 7-10/Mar/2010.

[47]

Mndez-Villafae R., Guerrero-Araque J.E., Gallego E., Lorente-Fillol A., Tore C., OrtizRodrguez J.M., Martnez-Blanco M.R., and Vega-Carrillo H.R. 241Am-Be source term simulation with Monte Carlo methods. In VI International Symposium on Radiation Physics. Zacatecas,
Zac., Mxico, 7-10/Mar/2010.

[48]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., and Mndez-Villafae R. Espectrometra y dosimetra neutrnicas con redes neuronales artificiales para un detector de 3He.
In 11 Seminario de Investigacin UAA. Aguascalientes, Ags., Mxico, 18-21/May/2010. Trabajo distinguido como la mejor ponencia.

[49]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., and Mndez-Villafae R. Comparacin del espectro de neutrones de un detector de 3He obtenido con dos cdigos de deconvolucin. In 11 Seminario de Investigacin UAA. Aguascalientes, Ags., Mxico, 18-21/May/2010.

[50]

Ortiz-Rodrguez J.M., Martnez-Blanco M.R., Vega-Carrillo H.R., Los Arcos-Merino J.M., and
Mndez-Villafae R. Clculo del error producido por las lecturas de conteo en la espectrometra
neutrnica por medio de esferas Bonner y la distribucin de poisson. In 11 Seminario de
Investigacin UAA. Aguascalientes, Ags., Mxico, 18-24/May/2010.

12/12

También podría gustarte

Neuronal Artificial
Aún no hay calificaciones
Neuronal Artificial
13 páginas
Memoria de Cálculo Invernadero para FVH
Aún no hay calificaciones
Memoria de Cálculo Invernadero para FVH
5 páginas
Las 5s
Aún no hay calificaciones
Las 5s
18 páginas
Reconocimiento de Placas Vehiculares Usando Redes Neuronales Artificiales
Aún no hay calificaciones
Reconocimiento de Placas Vehiculares Usando Redes Neuronales Artificiales
5 páginas
Informe de Hidropónicos
Aún no hay calificaciones
Informe de Hidropónicos
6 páginas
Preparación de Menas de Estaño
Aún no hay calificaciones
Preparación de Menas de Estaño
49 páginas
Ia Predicción de Variables en Planta Concentradora
Aún no hay calificaciones
Ia Predicción de Variables en Planta Concentradora
22 páginas
Desarrollar Red Neuronal para Android Studio
Aún no hay calificaciones
Desarrollar Red Neuronal para Android Studio
106 páginas
Administración de La Producción II, Indicadores de Gestión Enero 2021
Aún no hay calificaciones
Administración de La Producción II, Indicadores de Gestión Enero 2021
16 páginas
Remolque Basculante
Aún no hay calificaciones
Remolque Basculante
9 páginas
Geología de Cratones y Escudos
Aún no hay calificaciones
Geología de Cratones y Escudos
9 páginas
Criterios de Minado y Selección
Aún no hay calificaciones
Criterios de Minado y Selección
23 páginas
Evolución Escudo de Guayana MENDOZA - Desbloqueado
Aún no hay calificaciones
Evolución Escudo de Guayana MENDOZA - Desbloqueado
212 páginas
Forraje Hidropónico
Aún no hay calificaciones
Forraje Hidropónico
9 páginas
Hardware Con Inteligencia Artificial
Aún no hay calificaciones
Hardware Con Inteligencia Artificial
43 páginas
De 5s A 9s
Aún no hay calificaciones
De 5s A 9s
19 páginas
Diseño y Simulación de Espesadores
Aún no hay calificaciones
Diseño y Simulación de Espesadores
27 páginas
Ensayo de Kaoru Ishikawa
Aún no hay calificaciones
Ensayo de Kaoru Ishikawa
20 páginas
Explosivos en Construcción de Carreteras
100% (1)
Explosivos en Construcción de Carreteras
12 páginas
Redes Neuronales
Aún no hay calificaciones
Redes Neuronales
71 páginas
Condiciones Ambientales para Producir FVH
Aún no hay calificaciones
Condiciones Ambientales para Producir FVH
31 páginas
Corte y Relleno Ascendente
Aún no hay calificaciones
Corte y Relleno Ascendente
16 páginas
Linea de Tiempo Geologia
Aún no hay calificaciones
Linea de Tiempo Geologia
2 páginas
Las 5 y 9 S S
Aún no hay calificaciones
Las 5 y 9 S S
63 páginas
Corte y Relleno
Aún no hay calificaciones
Corte y Relleno
6 páginas
Estaño Miguel Chang
Aún no hay calificaciones
Estaño Miguel Chang
16 páginas
Motivación, Maslow, Clima Organizacional
Aún no hay calificaciones
Motivación, Maslow, Clima Organizacional
3 páginas
06 - Accesorios de Perforacion Rotativa
Aún no hay calificaciones
06 - Accesorios de Perforacion Rotativa
19 páginas
Mecanizacion de Invernaderos - 2
Aún no hay calificaciones
Mecanizacion de Invernaderos - 2
7 páginas
7 Herramientas de Calidad CD
Aún no hay calificaciones
7 Herramientas de Calidad CD
13 páginas
Redes Neuronales Presentacion
Aún no hay calificaciones
Redes Neuronales Presentacion
30 páginas
Reconocimiento de Voz y Electroacústica
Aún no hay calificaciones
Reconocimiento de Voz y Electroacústica
34 páginas
Como Programar Inteligencia Artificial
100% (1)
Como Programar Inteligencia Artificial
8 páginas
Tipos de Concentraciones
100% (1)
Tipos de Concentraciones
25 páginas
Sistemas Expertos Basados en Reglas
Aún no hay calificaciones
Sistemas Expertos Basados en Reglas
55 páginas
Herramientas de Inteligencia Artificial (IA) - PRESENTACIÓN
Aún no hay calificaciones
Herramientas de Inteligencia Artificial (IA) - PRESENTACIÓN
7 páginas
Sistemas Expertos
Aún no hay calificaciones
Sistemas Expertos
13 páginas
Clasificación Mecánica y Por Hidrociclones
Aún no hay calificaciones
Clasificación Mecánica y Por Hidrociclones
82 páginas
Informe Proyecto de Hackaton.
Aún no hay calificaciones
Informe Proyecto de Hackaton.
9 páginas
APUNTE 2 - TEORIA Curso Programación en Python Machine Learning Aplicado A Minería
Aún no hay calificaciones
APUNTE 2 - TEORIA Curso Programación en Python Machine Learning Aplicado A Minería
73 páginas
Ley de Corte Variable
Aún no hay calificaciones
Ley de Corte Variable
32 páginas
Mia Chanate PDF
Aún no hay calificaciones
Mia Chanate PDF
207 páginas
Review of Equipment For Mine Waste From The Conventional Thickener To The Deep Cone For Paste Production
100% (1)
Review of Equipment For Mine Waste From The Conventional Thickener To The Deep Cone For Paste Production
12 páginas
Tecnicas Modernas de Voladuras 3
Aún no hay calificaciones
Tecnicas Modernas de Voladuras 3
32 páginas
Levantamiento de Ventilacion de Una mina-ICM PDF
Aún no hay calificaciones
Levantamiento de Ventilacion de Una mina-ICM PDF
25 páginas
Mantenimiento Industrial Ruben Eduarte
Aún no hay calificaciones
Mantenimiento Industrial Ruben Eduarte
26 páginas
Transformación y Futuro de la Minería
0% (1)
Transformación y Futuro de la Minería
12 páginas
Diseño de Tronadura - Clase 2
Aún no hay calificaciones
Diseño de Tronadura - Clase 2
20 páginas
Prueba 1 Extracción, Carguío y Transporte
100% (1)
Prueba 1 Extracción, Carguío y Transporte
142 páginas
DV61 - Abr09 - Gestion de Costos Operaciones Subterranea
Aún no hay calificaciones
DV61 - Abr09 - Gestion de Costos Operaciones Subterranea
6 páginas
S9 - Logística Inversa e Indicadores de Gestión
Aún no hay calificaciones
S9 - Logística Inversa e Indicadores de Gestión
76 páginas
Metodologia 9s
Aún no hay calificaciones
Metodologia 9s
52 páginas
Detoxificacion
Aún no hay calificaciones
Detoxificacion
28 páginas
Cómo Aplicarlo - Bizagi
Aún no hay calificaciones
Cómo Aplicarlo - Bizagi
21 páginas
Presentación Clases N°2
Aún no hay calificaciones
Presentación Clases N°2
23 páginas
Métodos de Kriging en Geología
Aún no hay calificaciones
Métodos de Kriging en Geología
28 páginas
Análisis Granulométrico con ImageJ
Aún no hay calificaciones
Análisis Granulométrico con ImageJ
14 páginas
Antologia Redes Neuronales
Aún no hay calificaciones
Antologia Redes Neuronales
220 páginas
Aplicacion de Redes Neuronales Artificiales y Tecn PDF
Aún no hay calificaciones
Aplicacion de Redes Neuronales Artificiales y Tecn PDF
8 páginas
Molienda Redes Neuronalesj
Aún no hay calificaciones
Molienda Redes Neuronalesj
94 páginas
Une en 60269-1-A2 (1998)
Aún no hay calificaciones
Une en 60269-1-A2 (1998)
7 páginas
Une 21103-2-1 - Cei 60269-2-1 (1998)
Aún no hay calificaciones
Une 21103-2-1 - Cei 60269-2-1 (1998)
40 páginas
Une en 60269-4-A1 (1997)
Aún no hay calificaciones
Une en 60269-4-A1 (1997)
5 páginas
Une-En 15193 2008 (Ac 2010
Aún no hay calificaciones
Une-En 15193 2008 (Ac 2010
4 páginas
Une en 60269-3 (1996)
Aún no hay calificaciones
Une en 60269-3 (1996)
13 páginas
Capitulo X. Documentos de Referencia ENDESA
Aún no hay calificaciones
Capitulo X. Documentos de Referencia ENDESA
55 páginas
23-Un Paso Por Delante de La Bolsa (One Up On Wall Street) - Peter Lynch PDF
100% (3)
23-Un Paso Por Delante de La Bolsa (One Up On Wall Street) - Peter Lynch PDF
158 páginas
Curso de Homologación de Vehículos
Aún no hay calificaciones
Curso de Homologación de Vehículos
152 páginas
1 Guia Practica de Motores
Aún no hay calificaciones
1 Guia Practica de Motores
8 páginas
Test de Raven: Evaluación de Inteligencia
Aún no hay calificaciones
Test de Raven: Evaluación de Inteligencia
6 páginas
Educación Permanente y Educación Superior
Aún no hay calificaciones
Educación Permanente y Educación Superior
19 páginas
Morfología Léxica - La Formación de Palabras
Aún no hay calificaciones
Morfología Léxica - La Formación de Palabras
119 páginas
Exámenes de Epistemología
63% (32)
Exámenes de Epistemología
21 páginas
Trabajo Desarrollo Del Personal
Aún no hay calificaciones
Trabajo Desarrollo Del Personal
19 páginas
El Problema Del Conocimiento Obiols
Aún no hay calificaciones
El Problema Del Conocimiento Obiols
4 páginas
Manual de Capacitación Tutores
Aún no hay calificaciones
Manual de Capacitación Tutores
112 páginas
Cuestionario de Satisfacción Familiar - Docx Edith y Dayana 4B Aiep
Aún no hay calificaciones
Cuestionario de Satisfacción Familiar - Docx Edith y Dayana 4B Aiep
2 páginas
Posibilismo Ontológico
Aún no hay calificaciones
Posibilismo Ontológico
13 páginas
Tertulias Dialógicas Literarias: Guía Práctica
Aún no hay calificaciones
Tertulias Dialógicas Literarias: Guía Práctica
38 páginas
El Honrado Leñador
Aún no hay calificaciones
El Honrado Leñador
11 páginas
Conceptos de La Biblioteca Escolar
Aún no hay calificaciones
Conceptos de La Biblioteca Escolar
16 páginas
Conductivismo vs Constructivismo: Claves y Representantes
Aún no hay calificaciones
Conductivismo vs Constructivismo: Claves y Representantes
2 páginas
Tesis Dislexia
100% (1)
Tesis Dislexia
45 páginas
Teoría Sociocultural de Vygotsky
100% (1)
Teoría Sociocultural de Vygotsky
8 páginas
Tratamiento Del Dolor-Sugestion Hipnotica
33% (9)
Tratamiento Del Dolor-Sugestion Hipnotica
22 páginas
Ideas para Fortalecer La Educación A Distancia PDF
Aún no hay calificaciones
Ideas para Fortalecer La Educación A Distancia PDF
286 páginas
Comentario Resuelto Discurso 2parte 2012 PDF
100% (1)
Comentario Resuelto Discurso 2parte 2012 PDF
8 páginas
Teg - Rebeca Torrealba C.i.16596906
Aún no hay calificaciones
Teg - Rebeca Torrealba C.i.16596906
39 páginas
Clínica Psicoanalítica Con Adolescentes (Stella M. Firpo, Adriana Lasalle) PDF
100% (2)
Clínica Psicoanalítica Con Adolescentes (Stella M. Firpo, Adriana Lasalle) PDF
97 páginas
Desarrollo de La Perspectiva Teórica
Aún no hay calificaciones
Desarrollo de La Perspectiva Teórica
8 páginas
El Entorno Físico en Psicología de Grupos
Aún no hay calificaciones
El Entorno Físico en Psicología de Grupos
2 páginas
Epistemología de La Investigación Cualitativa - Martha J. Ruedas
Aún no hay calificaciones
Epistemología de La Investigación Cualitativa - Martha J. Ruedas
9 páginas
Teoría Interpretativa
Aún no hay calificaciones
Teoría Interpretativa
7 páginas
Filosofia Del Acontecimiento
100% (1)
Filosofia Del Acontecimiento
168 páginas
CEDIL-7S-CCastillo-EAI-EVALUACIÓN DIAGNÓSTICA
Aún no hay calificaciones
CEDIL-7S-CCastillo-EAI-EVALUACIÓN DIAGNÓSTICA
2 páginas
El Camino de Regreso Integración Colectiva
Aún no hay calificaciones
El Camino de Regreso Integración Colectiva
37 páginas
Producción Intelectual.
Aún no hay calificaciones
Producción Intelectual.
19 páginas
DEPRESIÓN
Aún no hay calificaciones
DEPRESIÓN
7 páginas
Informe de Lectura 3 - Heidegger
Aún no hay calificaciones
Informe de Lectura 3 - Heidegger
8 páginas

4 Inteligencia Artificial, UCO, TESIS

Cargado por

4 Inteligencia Artificial, UCO, TESIS

Cargado por

Universidad de Cordoba,

aplicaciones a la espectrometra y dosimetra neutronicas

Jose Manuel Ortiz Rodrguez

TITULO: Inteligencia artificial: aplicaciones a la espectrometra y dosimetra

TTULO DE LA TESIS: INTELIGENCIA ARTIFICIAL: APLICACIONES A LA

DOCTORANDO/A: JOS MANUEL ORTIZ RODRGUEZ

INFORME RAZONADO DEL/DE LOS DIRECTOR/ES DE LA TESIS

El trabajo desarrollado, y evidenciado a travs de las publicaciones realizadas, tiene el

emplean mtodos iterativos de solucin y tampoco requiere de tiempos largos de

Ortiz-Rodrguez, J.M.; Martnez-Blanco, M.R. and Vega-Carrillo, H.R. Evolutionary

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R.; Gallego, E. and Vega-Carrillo, H.R.

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009). A

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Por lo antes expuesto, se autoriza la presentacin de la tesis doctoral.

Crdoba, _9_ de _diciembre_ de __2010__

Fdo.: Dr. Hctor Ren Vega - Carrillo_

Fdo.: Jos Ignacio Benavides Bentez

aplicaciones a la espectrometra y dosimetra neutronicas

Todos los derechos reservados

Ninguna parte de e sta publicacion podra ser reproducida, almacenada o transmitida en

cualquier forma o medio electronico, mecanico,

TTULO DE LA TESIS: INTELIGENCIA ARTIFICIAL: APLICACIONES A LA

DOCTORANDO/A: JOS MANUEL ORTIZ RODRGUEZ

INFORME RAZONADO DEL/DE LOS DIRECTOR/ES DE LA TESIS

El trabajo desarrollado, y evidenciado a travs de las publicaciones realizadas, tiene el

emplean mtodos iterativos de solucin y tampoco requiere de tiempos largos de

Ortiz-Rodrguez, J.M.; Martnez-Blanco, M.R. and Vega-Carrillo, H.R. Evolutionary

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R.; Gallego, E. and Vega-Carrillo, H.R.

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009). A

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Ortiz-Rodriguez, J.M.; Martinez-Blanco, M.R. and Vega-Carrillo, H.R. (2009).

Por lo antes expuesto, se autoriza la presentacin de la tesis doctoral.

Crdoba, _9_ de _diciembre_ de __2010__

aplicaciones a la espectrometra y dosimetra neutronicas

no supe ni te pude dar, estando a tu lado acompan andote

observando mi evolucion, orientandome,

Agradezco al Dr. Jose Ignacio Benavides Bentez, Profesor titular de la Universidad de

to. Desde las primeras platicas

Agradezco al Dr. Roberto Mendez, Responsable del Laboratorio de Patrones Neutronicos,

por haber propiciado el hacer posible la realizacion de una estancia de investigacion

Agradezco a la Agencia Espanola

Agradezco a todos aquellos que no me es posible enumerar, amigos y familiares, por el

A todos ustedes: MUCHAS GRACIAS!

alta densidad, cuyos diametros

y los intervalos de clase de la energa de los

este sistema de ecuaciones, empleandose

el desarrollo de procesos complementarios. En las ultimas

es un problema trivial, ya que determinar un conjunto apropiado de valores de los parametros

1.2. Pregunta cientfica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.4. Objetivo General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1. Inteligencia Artificial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.1. Redes Neuronales Artificiales . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.2. Algoritmos Evolutivos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2. Espectrometra de Neutrones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2.1. Espectrometra de Neutrones con el SEEB . . . . . . . . . . . . . . . . . .

3.2. Pre-procesamiento de la informacion . . . . . . . . . . . . . . . . . . . . . . . . . .

3.3. Tecnicas de optimizacion de RNA . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Robusto de Redes Neuronales Artificiales . . . . . . . . . . . . . . .

3.3.2. Redes Evolucionadas por Algoritmos Evolutivos . . . . . . . . . . . . . . .

3.4.1. VMCNPex y generacion automatica

3.5. Optimizacion de las herramientas disenadas

3.6. Aplicacion de las herramientas en experimentos reales y comparacion con tecni

4.2. Captulo de libro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

2.2. Esquema basico

2.3. De la neurona biologica a la neurona artificial . . . . . . . . . . . . . . . . . . . .

2.4. modelo de una neurona artificial . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.5. Esquema de una Red neuronal de una capa con S neuronas . . . . . . . . . . . . .

2.6. Esquema general de una RNA de 3 capas . . . . . . . . . . . . . . . . . . . . . . .

2.7. Posible clasificacion de las redes neuronales . . . . . . . . . . . . . . . . . . . . . .

2.8. Entrenamiento de una Red Neuronal . . . . . . . . . . . . . . . . . . . . . . . . . .

2.9. Estructura celular biologica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.10.Analogas en los algoritmos geneticos I . . . . . . . . . . . . . . . . . . . . . . . . .

2.11.Analogas en los algoritmos geneticos II . . . . . . . . . . . . . . . . . . . . . . . .

2.12.Analogas en los algoritmos geneticos III . . . . . . . . . . . . . . . . . . . . . . . .

2.13.Modelo de fitness que ilustra la forma de operar de un AE . . . . . . . . . . . . .

Crdoba, _9_ de _diciembre_ de 2010

Crdoba, _9_ de _diciembre_ de 2010