Genetic Algorithms and Cellular Automata: A New

Architecture for Traffic Light Cycles Optimization

Javier J. S h c h e z Manuel Gal& Enrique Ruhio
Centro de Innovacidn para Centro de Innovacidn para Centro de Innovacidn para
la Sociedad de la Informacidn la Sociedad de la Informacidn la Sociedad de la Informacidn
(C.I.C.E.I.) (C.I.C.E.I.) (C.I.C.E.I.)
University of Las Palmas de G. C. University of Las Palmas de G. C. University of Las Palmas de G. C.
35017-Las Palmas de Gran Canaria, 35017-Las Palmas de Gran Canaria, 35017-Las Palmas de Gran Canaria,
Spain Spain Spain
Email: jsanchez@polaris.ulpgc.es Email: manolo@cicei.com Email: rubio@cicei.com

Abstract- We present a new architecture for the optimization The rest of this article is organized as follows. In the next
of ’lYafEc Light Cycles in a ’lkaIlic Network. The model is based subsection we comment on some other group efforts regarding
on three hasic design items: The use of Genetic Algorithms as an traffic optimization. In section I-B we explain some concepts
optimization technique, the use of Cellular Automata Simulators
within the Evaluation Function, and the use of a Beowulf Cluster related to traffic simulation. In section I1 we formulate the
as parallel execution environment for this architecture. We also problem. Section 111 describes our architecture in three steps,
present some tests to demonstrate that this new architecture is namely Genetic Algorithm description, Traffic Simulator de-
suitable for the problem it solves. scription and Beowulf Cluster description. Section I V c o n -
ments some of the test results we have obtained. Finally,
I. INTRODUCTION section V includes our conclusions and some future work
ideas.
Nowadays traffic management is a first priority problem for
many cities. Often extending traffic infrastructures is not viable A. Other Groups Efforts
and it is a very expensive solution. Moreover, environment Many research groups have dedicated their efforts to im-
impact is a very important matter we cannot forget. prove the use of traffic infrastructures. Some groups wclrk
It would be more convenient to optimize the existing infras- on the optimization of traffic light cycles at a stand-alone
tructures in order to get the best service from them, before new intersection. For example in [ l ] every intersection is opti-
ones are constructed. For this reason, many research groups mized considering only local information. Moreover, it can
around the world are working on optimization methodologies be adapted to short and long time traffic fluctuations.
for traffic. Other groups work on the optimization of several intersec-
Traditionally, the traffic management has been approached tions at the same time. An example is [2]. In this work an
using the Trial-And-Error method: an expertise person or “ad hoc” architecture is used to optimize a 9 intersections
team decided on traffic parameters and depending on the traffic network. It uses Genetic Algorithms as an optimization
resulting traffic behavior some feedback corrections occurred. technique running on a single machine. The CORSIM model is
This philosophy hasn’t changed so much in past decades, used within the evaluation function of the Genetic Algorithm.
except for the use of simulators instead of real traffic tests In this work scalability is not solved. Authors recognize that
as feedback source. Recently some micro-simulators - based it is a customized non scalable system.
on the vision of traffic as a collection of independent vehicles The offset time between two traffic lights is optimized with
- have proved to be very accurate. Artificial Neural Networks ( A ” s ) by [3]. The ANN is trained
Although the use of these tools has permitted the execution with the data provided from a simulator built up for this
of many different simulations in relative short time,the method purpose. The ANN produces estimations of the length of queue
still depends on the engineer experience and ”an”, and more- of vehicles stopped in front of a traffic light. It allows the
over it can’t ensure that the whole search space is covered. calculation of the optimal offset between two consecutive r.ets
Traffic optimization comprises a set of different problems, of traffic lights.
from which one of the most relevant ones is the traffic light In [4] the stop and go traffic waves are minimized using
cycles optimization. Traffic light cycles optimization is a N p - an integral functional. They face it as a non standard prob-
Complete problem. Considering that in these NP-complete lem of calculus of variations. Given a system of hyperbolic
problems a deterministic optimizing method is not known, we conservation laws, they introduce an integral functional.
find suitable to use non-deterministic optimizing methods like Currently we optimize the traffic light cycles and the
Genetic Algorithms. maximum speed allowed in a set of interconnected streets.

0-7803-85 15-2/04/$20.0002004 lEEE 1668

In this work we present only the results of traffic light cycles In [IO] we can find a well described list of microscopic
optimization. We use a Cluster Beowulf as a cheap, scalable models and a comparative study of them. Although conclu-
and flexible MIMD parallel computer. sions are not definitive, this work seems to demonstrate that
models using less parameters have better performance.
B. Traf/ic Simulation We have developed a traffic model based on the SK model
([I I]) and the SchCh model ([ 121). The SchCh model is a com-
Traffic Simulation is known to be a very difficult task. There
bination of a highway traffic model - Nagel-Schreckenberg
are two different sorts of traffic models. The first one is the
[13] - and a very simple city traffic model - Biham-
macroscopic model set. They are based on Fluid Dynamics Middleton-Levine [14]. The SK model adds the “smooth
since they consider traffic flow as a continuous fluid. On the
braking”. We decided to base our model in the SK model
other hand, we have microscopic approaches. In them, traffic
due to its better results for all the tests shown in [IO].
is considered as a set of discrete particles following some
rules. In the last decade there is a common belief about the The Nagel-Schreckenberg model simulates highway traffic.
better performance of Microscopic approaches to do Traffic It consists of an one-dimensional lattice of cells that can
Modeling. One such widely used approach is the Cellular accommodate no more than one vehicle at a time. Each vehicle
Automata Model. is characterized by a maximum velocity and a randomization
Scientific literature is plenty of macroscopic approaches parameter. It considers the number of free cells in front
for traffic modeling. In the 50s appeared some “first order” and accelerates when possible to the maximum velocity or
continuum theories of highway traffic. In the 70s and later decelerates when necessary. An extra random parameter makes
some other “second order” models were developed in order the vehicle to decelerate even if it is not necessary
to correct the formers’ deficiencies. References [ 5 ] , [61, and In the Biham-Middleton-Levine model a square lattice mod-
[7] may illustrate some of these models. In [5] a continuous els the traffic network. Each cell represents a crossing of east-
traffic model is proposed taking into account the safe distance bound and north-bound streets. It can be empty or occupied
each driver keeps. Moreover it extends the Navier-Stokes- by a vehicle moving to the north or to the east. At odd time
like equations for the velocity variance of vehicles. In [6] steps only the east-bound vehicles are updated, and at the even
the nonlinear theory of the cluster effect in a traffic flow, time steps the north-bound vehicles are updated.
i.e., the effect of the appearance of a region of high density The SchCh model combines the Biham-Middleton-Levine
and low average velocity of vehicles is presented. In [7] model with the Nagel-Schreckenberg model rules for ve-
it was presented a very well known free way simulation hicle dynamics. The SK model is an extension of Nagel-
model. It formulated a variant of the equilibrium speed-density Schreckenberg. It avoids discontinuous velocity changes.
hypothesis that overcame the spontaneous lockup problem by
In [I51 it is demonstrated that the traffic light cycles have a
adding a “look-ahead” term to the speed equation. strong influence in traffic flow results. This is the reason why
However, for example in [8] “second order” models are we decided to deal with this problem.
questioned due to some serious problems, i.e. a negative flow
predictions and negative speeds under certain conditions.
Nowadays the microscopic simulators are widely used.
Their main drawback is that they have a lower performance
than the macroscopic ones, except in the case of the Cellular
Automata. Cellular Automata and macroscopic simulators
have similar computing times.
The Cellular Automata Simulators are based on the Cellular
Automata Theory developed by John Von Neumann [9] at the
end of the forties at the Logic of Computers Group of the
University of Michigan. In this theory vehicles - in the traffic
simulation case - are considered as discrete unidimensional
entities. The streets are sampled into a set of points. There
can be only one vehicle at each point. It establishes the
rules over them and let them circulate freely. In the Cellular
Automata model not only space is sampled into a set of
points, but also time and speed are sampled too. Time becomes
iterations. It sets a relationship between time and iterations,
e. g. l s e c a d E literatim. Consequently, speed tums into
”points over iterations”. If we sample a point every 2m and
we have a vehicle moving at I O d s (36Km/h) in our Simulatol
Fig. 1. Initial tralfic lopology
it will be represented as a vehicle moving at a 5 points per
iteration speed.

1669
 11. PROBLEM DEFINITION 3 ) Evaluation and Fitness: For the evaluation we use
the mean time within the system. This is the elapsed time
:We start from a traffic network like the one in figure I .
(iterations) since a new vehicle arrives to the network until it
It consists of five two-lane streets. The symbol 'S' means a
leaves. We do believe that this simple and intuitive parameter
traffic light. Circulation sense is indicated with arrows. For
gives an appropriate picture of the network performance.
*is network we try to find the optimal cycle of traffic light
Although we have tested other parameters such as traffic
tombinations during a fixed period - 'PERIODSIZE'. This
homogeneity, queues' lengths, etc., the mean time within the
cycle is repeated indefinitely.
network seems to he the hest one, also producing quite robust
The restrictions are the following: results. We do not discard the employment of other statistical
1 ) There can only be a green traffic light by intersection values, whenever we can register improvements.
every cycle step. Our optimization problem is a minimization problem. This
2) Too short traffic light transitions are forbidden. Every is the reason why we calculated Fitness simply as the inverse
cycle step comprises a constant number of simulation of the evaluation value - Mean time at the system.
iterations - 'Traffic light granularity'.
3) In our work we only consider two traffic light states B. Cellular Automata Simulator
namely green and red. In our model these are the rules applied to all the vehicles.
4) The input data for our optimization is set off-line, and 1) A vehicle ought to accelerate up to the maximum speed
is based on statistical information. allowed. If it has no obstacle in its way (other vehicle,
or a red traffic light), it will accelerate at a pace of 1
111. NEW ARCHITECTURE DESCRIPTION point per itemion, every iteration.
A. Generic Algorithm Description 2) If a vehicle can reach an occupied position, it will reduce
its speed and will occupy the free position just behind
I ) Chromosome Codification: In this work we optimize the
the preceding.
traffic light cycles for all the intersections of a traffic network.
3) If a vehicle has a red traffic light in front of, it will stop
Every cycle is represented by a word of PERIODSIZE
at it.
integers. In every integer it is coded with two bits which traffic
4) Smooth Braking: Once the vehicle position is updated,
light is open at every cycle step. In figure 2 we show our
the vehicle speed is also updated. To do this, the number
chromosome.
of free positions from current position forward is taken
Period Sire ( 6 4 , 1 2 8 ...1 ./. P e r i o d Size
> into account.
)212[01@111213111 ... lOlllOlolllol2lal2loI ... 1211 .. . 5) Multi-lane: When a vehicle in intending to move on,
lniersecrio" @ TnrerSection 1 or to calculate the number of free positions forward on
< Perrod Size
> its way, it is permitted to consider positions on other
10 10 00 00 ff . . . 1,q*,011131310121 ... I2U parallel lanes.
-&3 InterreCtiOn N - 1 Multi-lane moving is achieved by means of storing the set
of possible next positions for every point in the scheme and
Fig. 2. Traffic Light S l l e Chromosome for every input-output couple. For example, in figure 3 there
is a vehicle in the path from point 24 to point 13. If it is in
2 ) Selection, Crossover and Mutation Operators: We have point 25 it may move to points 2 or 3.
chosen an Elitist Selection Strategy. It means that at every In [16] a more complex multi-lane traffic set of rules is
generation a little group of individuals - the best two indi- presented. However, our multi-lane scheme is simpler. M o r e
viduals in our case - is cloned to the next generation. The over it has a very iuteresting feature. Our scheme permits a
rest of the next generation is created by means of crossover vehicle to change to other lane twice of more times just in
of their "parents" and mutation - those chosen by mutation the same updating step. For example, if a vehicle has a speed
probability. of 2 points per iteration, it may change to other lane twice in
Some other selection strategies like Tournament Selec- the same simulator iteration. We believe that this is a more
tion and Probabilistic Tournament Selection have been tried. realistic behavior.
However, for our problem they seem to provoke premature Our Cellular Automata simulator is not stochastic. So far,
convergence. we avoid stochasticity. The position updating is done by
A standard two points crossover is used. We employ a means of a recursive function. It explores all the network
variable mutation probability. This is because we believe that a dependencies and orders the updating sequence, so deadlocks
high initial mutation probability would diversify the population are avoided. Hence, in every simulator iteration all the vehicles
and avoid local Minna. When a mutation occurs we proceed have the possibility of movement - if there is a free position
as follows. In the chosen individual we randomly chose the or the car in front moves too.
mutating gene position from all genes - and all intersections. Initial experiments showed that our Cellular Automata Sim-
Then we overwrite its value by a random value inside the rank ulator had a stochastic behavior. It is a very undesirable char-
of traffic lights at this intersection, i.e. [0,3]. acteristic, because we need that our evaluation function retunis

1670
always approximately the same value for an individual in order Nodes are connected through a Gigabit Ethernet Backbone.
to guide the Genetic Algorithm up to the better solution. In a Every node has the same hardware, except the master node
second version of our simulator we have given it more realism, having an extra Gigabit Ethernet network card for "out world"
letting the vehicles more flexibility for movement in order to connection. The price of this hardware nowadays is about
reduce the variance of results. This happened but it was not €5500.
enough. This is the reason why we finally decided to eliminate 2) Sofrware: Every node has installed Red Hat 9 on it -
the random behavior for good. Kernel 2.4.20-28.9, glibc ver. 2.3.2 and gcc ver. 3.3.2. It was
We realize that traffic flow is an intrinsic stochastic process. also necessary for parallel programming the installation of
However, we need to obtain deterministic results from our LAM/MPI (LAM 6.5.8, MPI 2). For genetic programming,
simulator in order to guide the Genetic Algorithm. We do many Genetic Algorithms free libraries are available. How-
believe that our simulator, even if its behavior is not stochastic, ever, in this work all the genetic algorithm architecture was
gives us a very accurate idea of how well does a chromosome developed in just plain ANSI C, only with the MPI parallel
work, and permits us to compare it to others by means of a routines help.
well defined fitness function. In our application there are two kinds of processes, namely
master and slave process. There is only one master process
running on each test. In every generation it sends the chromo-
somes (MPISend) to slave processes, receives the evaluation
results (MPIRecv) and creates the next population. Slave
processes are inside an endless loop, waiting to receive a new
chromosome (MPIRecv). Then they evaluate it and send the
evaluation result (MPISend). The chromosome consists of a
64 bytes array when PERIODSIZE - The number of states
in a traffic light cycle - is set to 64 steps, and 128 when it
is set to 128 steps. The evaluation value is a double type (8
bytes)

TABLE I
M.E.T., SPEED-UP, A N D B VARYING THE APPLICATION
SCHEME

- __ __ __
slave processes dishbution __ s
M.E.T. __ B(%) __S!A
Serial (N = I) 1052.12 - __- 2.20551
N = 2 (I slavelnode) 533.65 1.9716 38.4012 1.56598
N = 3 (I slavdnode) 363.69 2.8929 18.2094 4.98604
N = 4 (1 slavelnode) 272.21 3.8651 7.3405 0.61670
N = 5 ( I slavdnode) 220.30 4.7759 1.1731 0.39385
N = 5 (2 slavdmde) 324.21 3.2452 13.5186 1.08224
C. Beowulf Cluster Description N = 5 (5 slavelnode) 386.89 2.7194 ?0.9655 0.67207
Each time applications require more and more computing N = 5 (IO slaveslnode) 416.33 2.5271 ?4.4634 5.31319
power. In the past, powerful systems were very expensive N = 5 (20 slaveslnode) 499.33 2.1046 34.3929 34.70410
and thus prohibitive for most potential users. This situation N = 5 (40 slavednode) 530.62
__ 1.9828
__ 38.0421
__ 55.45 18 1
___
has changed in the last years, since in 1994 T. Sterling and
D. Becker developed a "distributed computing cluster" over
Linux Operative System and using "Of The Shelf' hardware
IV. TESTSPERFORMED
components. This new flexible, reusable and massively parallel
sort of MIMD has a very good price/performance relationship, We present three tests in this section. The first one is a
and is very suitable for intrinsic parallel tasks like the Genetic Speed-up study in order to demonstrate that this is a useful
Programming ones. architecture for parallelizing optimization tasks. In section IV-
The Architecture of our system is based on a five node B we present one example optimization with some partial
Cluster Beowulf, due to its very interesting pricelperformance results and a fitness evolution picture. At last in section IV-
relationship, and the possibility of employing Open Software C there is a comparison between the results of an optimized
on it. On the other hand, this is a very scalable MIMD network and the results of other two non optimized systems.
computer, a very desirable feature in order to solve all sort The traffic network of our tests is shown in figure 1. In
of traffic problems. figure 3 we show the sampled scheme. We have considered
I) Hardware: Every cluster node consists of a Pentium IV approximately 7m. between two points - the minimal dis-
processor at 3.06 GHz with 1 GB DDR RAM and 80GB HDD. tance required in a traffic jam. In figure 3 we have represented

1671
every traffic light with a rectangle, and inputs and outputs with
triangles.
For all the tests performed there are some fixed values. The
population size is 300 individuals. Every test goes on during
300 generations. We have used a variable mutation probability.
It starts with a very high value (0.75) and decreases by a
factor of 0.982. The relationship between time and iterations
is literatim = lseconds, more or less the human response
time. And every simulation runs during 1500 iterations. For
the tests in sections IV-A and IV-B PERIODSIZE is set Fig. 5. Intersection 0 panial results
to 64 and Traffic light granularity is set to 5 iterations. For
test in sections IV-B and IV-C we chose the best Speed-up
application scheme we have found so far - 1 master and 5 is 1.1731. It demonstrates the very high parallelism of genetic
slaves processes. algorithm.
When we increase the number of slave processes at every
A. Parallel Speed-up Study node the performance - execution time - decreases. We
In table I we represent the results. 10 executions have been believe that only one slave process per node permits the best
used to calculate every mean value. In the first column we have computation and communication overlapping. When there axe
the slave processes distribution. The input named Serial means more than one slave per node there is a resource competition at
only one master and one slave processes running over the same every node that provokes not only a linear increase of slaves
cluster node. It always runs one slave process per node except processing time, but also some context changing overhead.
in the last five tests. In the second column the mean execution This seems to be the reason of the performance worsening.
time (M.E.T.) for every application scheme is presented. The B. Test Example Optimization
third and fourth columns represent the Speed-up (S) and
the serial percentage of our algorithm (B), both calculated 30 executions were performed for every PERIODSIZE and
following Amdahl's law (eq. 1). N means the number of granularity couple. We made tests with values of 64 and
processors. Finally the standard deviation of time results is 128 steps for PERIODSIZE and 5 , 10, 20, 30, 40, 50 and
shown. 60 iterations for granularity. We present the results from the
best test. It was achieved for PERIODSIZE 728 steps, and
granularity 3 iterations. For this case the mean execution time
is 1420.52s. We present the results from the best test. In figure
4 it is presented the best individual fitness evolution, all tests
mean fitness evolution, and the Standard Deviation of them.
0.095 -
The best fitness had an increase of 41.99%. The mean fitness
had an increase of 39.60%.In figure 5 we displayed a fragment
(16 steps) of traffic light cycles of one intersection. A black
box means red light and a white box means green light. Evi:ry
cycle step is equivalent to 5 simulation iterations - Traffic
Light granularity - and consequently to 5 seconds.
The conclusion of this results is that we have a suitable
genetic optimization in 300 generations.

3 C. Optimiultion Comparison
We have carried out several comparative studies to test how
2.5
good are the optimized results in comparison to other simpler
2- approaches. To do so we have carried out a set of tests with
two different simulators. In the first simulator CRanSim) the
traffic light cycles follow a random sequence.
In the second simulator (FixSim) there is a fixed green time
Fig. 4. Best Fitness and Average Fitness Evolution. and Standard Deviation in a fixed order for traffic lights. Every traffic light has the
same green time as the rest.
From table I two relevant discussion subjects emerge. First, For these two simulators and our system we have canied
one can notice a performance improvement as one increases out two sets of tests, one setting the PERIODSIZE value to
the number of nodes sharing the work. The best situation is 64 steps (6), and one setting if to 128 steps (7). We do this to
obtained when only one slave process is running at every observe if there is any difference in results varying the traffic
node. For this situation there is a Speed-up of 4.7759 and B light cycle length.

1672
 275
2s
225
2cm
r,
8s
ZL:
175
2 . 180
-* -
~

f,
.- 12s
a% 100
55
3
75
E 50
25
0
5 10 20 30 40 50 60 5 10 20 30 40 SO 60
Traffic Liphu Granuladty Traffic LlghU Granulamv

Fig. 6 . Mean Times a1 Network (s) for PERIODSIZE=M Fig. 7. Mean Times at Network (s) for PERIODSIZE=128

25
22.5
20

3 c 17.5
I5

P 5
2.5
0
5 10 20 30 40 50 60 5 10 20 30 40 SO 60
Traffic LlghfS Granularity TrafficLighu Granulamv

Fig. 8. Mean Times a1 Network (5) for PERIODSIZE=M Fig. 9, Mean Times at Network (s) for PERIODSIZE=IZ8

Additionally, we present other optimization results. This increase the traffic lights granularity
simulator (LISim) comprises, a strong feedback and local In second place it seems that there is not a profound
intelligence system by means of a dynamic updating traffic difference between results with a PERIODSIZE of 64 or
lights state d e . The next green is decided depending on 128 steps. The greater difference is a 6.5 % for the case of
the number of occupied points behind the traffic lights at an granularity equal to 60 with our system.
intersection. The traffic light with the bigger number is chosen. In third place, there is a strong improvement in our system
If there is more than one traffic light with the same number from RanSim and FixSim. There is a mean decrease of a
of vehicles enqueued the next green is chosen randomly. 83.82% comparing with RanSim and 56.74% comparing with
In figure 6 we present the mean time comparison for FixSim.
RanSim, FixSim and our optimized network setting the PE- Finally, from figures 8 and 9 we get a mean improvement
RIOD-SIZE value to 64 steps. In figure 7 it is shown the of 42.78% in LISim against our system.
mean time setting the PERIODSIZE value to 128 steps. 10000
simulations were used to calculate every mean time of FixSim v. CONCLUSIONS AND FUTURE
WORK
and RanSim for every granularity value. Every result of our We propose a new architecture, based on the combination
system is the mean value of 30 executions. of Genetic Algorithms over a MIMD computer - a Beowulf
In figure 8 we present the comparison for LISim and our Cluster - using a Cellular Automata simulator within the
optimized network with PERIODSIZE equal to 64 steps. evaluation function.
In figure 9 it is shown the same comparison setting the Up to date traffic network optimization has been faced using
PERIOD-SIZE value to 128 steps. Again 10000 simulations the Trial-And-Error method. This method cannot ensure that
were used to calculate every mean time of LISim and 30 the whole searching space is covered. We propose a new
executions for our system. method, a non deterministic optimization one to solve this
Some conclusions may be obtained observing these figures. task. In this work we optimize traffic light cycles. However,
In all cases there is a progressive worsening in results as we every traffic parameter - or set of parameters - may be

1673
 optimized following a similar procedure, and every scale of REFERENCES
traffic network may also be optimized - scaling up hardware [I1 Vogel. A.. Goerick, C., von Seelen, W.: Evolutionary Algorilhms for
too. Optimizing Trafh Signal Operation. Proceedings of ESIT 2w0, Aachen.
Test IV-B demonstrates that this is a valid architecture for Gemany (ZWO) 83-91
[21 Rouphail, N., Park B., Sacks.J.: Direct Signal Timing Optimization:
optimizing traffic light cycles. Moreover it is important to note Strategy Development and Resulls. In XI Pan American Conference in
that all intersections are optimized at the same time. We do Tr2flic and Transportation Engineering, Gramada. Brazil (ZWO)
believe that this sort of global optimization will produce better I31 L6pez. S., Hemhdez. P.,Hemhdez, A., and Garcia, M.: Anificial Neural
Networks as useful tools for the Optimization of the relative orset betwtrn
results than a stand-alone intersection optimization, such as two consecutive sets of traffic lights LNCS. Foundations and Twls for
[I]. Neural Modeling. Springer-Verlag. (1999) 795-804.
The tests performed in IV-C are very interesting. It is 141 Colombo, R. M., Groli, A.: Minimising Stop Go Waves to Optimise
demonstrated that our optimized traffic light cycles improve ..
Traffic Flow. A d . Math. Letten. To amcar.
[51 Helbing. D.: An improved fluid dyn&bal model for vehicular traffic.
the results of other simpler approaches -FixSim and RanSim. Physical Review. E51, (1995) 3 1 6 4
Secondly one may observe that results from LISim are [6] Kemer, B. S . and Konhsuser, P.: Smcture and Parameters of Clusten: in
Traflic Flow Physical Review E.5054,(1994)
much better than results from our architecture. LISim has a %bibitemKuhne.
great advantage over our system. It is that it employs strong [7l Payne, H. 1.: FreRo: A macroscopic simulation model of freeway trallic.
feedback information for the next green decision. Transportation Research Record. 722:68, (1979).
I81 Daganzo.C.F.: Requiem for second order fluid approximations of mfiic
Frequently it is not possible to implement such a feedback Row. Transportation Research B, (1995)
system - like the one used by LISim - in a real life [91 von Neumann.J.: The General and Logical Theary of Automata. John yon
situation. The most of nowadays traffic networks are not yet Neumann-Collected Works, Volume V, A. H. Taub (ed.).Macmillan,
New York, (1963) 288-328.
prepared to give this sort of real-time information. So far the [IO] Brockfeld, E.; Khne. R. D.; Wagner, P.: Towards Benchmarking Mi-
only information one can get in those systems is statistical croscopic Traffic Flow Models Networks for Mobility, Intemational
information. Symposium. Volume I, (2W2) 321-331.
[ I l l Krauss. S., Wagner, P.. and Gawran. C.: Metaslable states in a micm-
However we do believe that the hest way to deal with traffic scopic model of uaffic flaw, Phys. Rev. E55 5597 - 5605 (1997)
networks optimization goes through the implementation of ~. Schadschneider, A ;Chawdhurv, D : Bmkfeld. E ; Klauck K : Santen,
1121
systems using feedbackinfomati& to co& rt. its optimiza. L ;zittanz, I: A new cellular automata model for city u a f f i c . ' ~ r a f l i c and
Granular Flow '99: Social. Trail% and Granular Dynamics". Springer,
tions in real-time. This is the next step we are taking in our Berlin (1999,
, ~~,
research. [I31 Nagel. K.. Schreckenberg. M.: A Cellular Automaton Model for Freeway
Finally we want to note that our architecture has one ad. TrafliC. Joumal de Physique I France, 332 (1992) 2221-2229
[I41 Biham, 0.. Middleton. A. A., Levine, D.: Phys. Rev. A 4 6 6124 (1992)
ditional benefit' It permits to multiply performance just [I51 Brockfeld. E., Barlovic, R., Schadschneider, A.. Schreckenberg, M.:
scaling up hardware. Test results in section N - A demonstrate Optimizing Traffic Lights in a Cellular Automaton Model for City ~ d k
this. This high parallelism is a very desirable feature for big http:~/aniv.orglPs/cond-ma1/0107056
[ 161 Wagner, P., Nagel, K. and Wolf. D. E.: Realistic Multi-Lane Traffic Rules
networks or real time optimizations. for Cellular Automata. Los Alamos National Laboratory repon LA-IJR-
96-2586, (1996)

1674