JPH04184668A - Correcting method for tutor data of neural network - Google Patents

Abstract

PURPOSE:To reduce the tutor data and shorten the learning time by composing data from data which is short in distance based upon a specific distance function and reconstituting new tutor data when the tutor data is added to specific tutor data. CONSTITUTION:When an input value (room temperature and temperature variation) is supplied at the time of air conditioner operation, an air conditioner output value (air blow temperature) which is judged by the neural network 2 to be the best value is outputted. A manual operation value is written in a register 4 successively as the new tutor data together with the input value. When the air conditioner operation ends, several tutor data in a RAM 1 and (m) tutor data in the register 4 are divided by a clustering means 5 into (n) groups and a representative point is determined. The three tutor data on the room temperature, temperature variation, and air blow temperature are put together from data which is short in distance based upon the specific distance function and their mean is used as new data. This operation is repeated to decrease (m+n) data to (n) data and the network 2 is made made to relearn.

Description

DETAILED DESCRIPTION OF THE INVENTION (A) Field of Industrial Application The present invention relates to a method for modifying training data for a neural network system that has a hierarchical structure and changes coupling coefficients through learning.

1) Conventional technology Learning One of the neural network systems that are currently widely used to perform input/output conversion according to data is the pack propagation method (hereinafter abbreviated as BP method).

).

The BP method targets networks with multilayer structures without feedback, and changes the strength of connections so as to minimize the square error between the output and the training data.

In this way, if the P method is trained by giving training data,
It has the characteristic of being able to learn any type of input/output relationship.

However, conversely, operation in areas where training data is not provided is not guaranteed at all. For this reason, exchanging a part of the teaching data with another teaching data may create an area where no teaching data is given, and conventionally the method has been to add new data to the teaching data.

(c) Problems to be Solved by the Invention However, in the conventional method, the number of teacher data increases rapidly by adding new teacher data, and the learning time also increases accordingly. Moreover, increasing the number of teacher signals requires securing memory for the teacher data.
It has always been necessary to impose restrictions on the addition of teaching data.

(d) Means for solving the problem The present invention has been developed in view of the above points, and includes:
When adding teacher data to preset teacher data,
The amount of training data is reduced by combining data with smaller distances based on a predetermined distance function and reconstructing new training data.

(e) The action training data is usually selected to represent the entire input/output space. If there are m pieces of initially given teacher data and n new teacher signals are added thereto, there will be a total of n pieces of teacher signals. Then, if we combine the ones that are closest to each other and find m new teaching data -, these m teaching data will be the new and old m
+n pieces of teacher data will be represented.

(f) Example First, consider installing a neural network system in a device such as an air conditioner to perform control according to the usage environment. A standard environment is trained in advance using standard teacher data, and if the user does not like this standard operation, the user operates the switch manually. As a result, this manual operation operates preferentially, and the input/output relationship at that time is appropriately sampled and stored.

After the power is turned off, this manual input/output relationship is used as new teacher data, clustering is performed together with the old teacher data, the teacher signals are returned to the original number, and learning is performed using the teacher signals.

FIG. 1 shows a teacher data correction device for implementing the present invention, which is usually implemented by an information processing device such as a computer. For example, in an air conditioner, n pieces of information necessary for operating the air conditioner, such as room temperature, room temperature change, and corresponding air blowing temperature, are stored in RAM 1 as initial teacher data. The neural network 2 is normally trained based on this teacher data.

When operating the air conditioner, when input values (room temperature, change in room temperature) are given, an air conditioner output value (blow temperature) determined by the neural network 2 to be optimal is output. If the user is not satisfied with this output, he or she performs manual operation using operation keys or the like. At this time, the output value by manual operation is preferentially selected by the selection means 3 and the air conditioner is operated. Further, this manual operation value becomes new teacher data as an output value together with input values such as room temperature and changes in room temperature, and is sequentially written into the register 4.

When this air conditioner operation is completed, the n pieces of teacher data in the RAMI and the m pieces of teacher data in the register 4 are divided into n groups by the clustering means 5, and each cluster has a representative point (representative teacher data). can be determined.

When using room temperature, room temperature change, and air blowing temperature as elements of training data, such as in the case of an air conditioner, first consider making the units dimensionless during clustering. Also, since room temperature changes and room temperature values vary depending on the unit, multiply them by an appropriate coefficient. Under these conditions, the room temperature Tl ('C) and the room temperature change T2 (
"C)" to derive KITl and K2T2. Similarly, 3T3 is derived from the air conditioner air temperature T3. For simplicity, set X = KIT1, Y = 2T2, Z = 3T3, and calculate the i-th r-ta. The Euclidean distance between (Xi, Yi, Zi) and the third data (Xi, Yi, Zi) is d, = [(Xi-Xj)"+ (Yi-Yj)"+(
It is expressed as 178 months. Then, the training data is integrated in order of decreasing distance, and the average of the integrated data is used as new data. This process is repeated to convert m10n pieces of data into n pieces of data.

The neural network 2 is trained using the milled data in this way.

In addition, in this example, the Euclidean distance is used in a hierarchical manner,
Although we explained how to take the group average as the representative point (representative teacher data), there are no restrictions on the clustering method, and standard Euclidean distance, Mahalanopis distance, Minkowski distance, etc. may be used, and how to determine the representative point. Also, the shortest distance method, the longest distance method, the centroid method, the quadratic method, etc. may be used.

(G) Effects of the Invention As described above, in the present invention, even if teacher data corresponding to the usage environment is added to the standard teacher data, the characteristics of all the teacher data are utilized to reduce the total number of teacher data. This reduces the neural network's maximum learning time.

[Brief explanation of the drawing]

FIG. 1 is a block diagram for implementing a method for correcting training data in a neural network according to the present invention. 1...RAM, 2...Neural network, 3
. . . selection means, 4 . . . register, 5 . . . clustering means.

Claims (2)

[Claims] (1) When adding teacher data to preset teacher data, new teacher data is reconstructed by combining data from those with a small distance based on a predetermined distance function,
A method for correcting training data in a neural network system characterized by reducing training data. (2) A method for correcting training data in a neural network system according to claim 1, characterized in that a Euclidean distance function is used as the distance function.