JPH08129542A - Parameter updating device for neural network - Google Patents

Abstract

PURPOSE: To easily mount a genetic algorithm for the update of parameters as the coupling coefficient and threshold value of the neural network by substituting an optional random number for the value of a gene seat at the time of mutation processing. CONSTITUTION: Coupling coefficients W and threshold values θ of neurons 3 constituting the neural network are regarded as constituent elements of a gene 5. Then, one gene 5 is constituted by regarding all coupling coefficients W and threshold values θ in the neural network as the values of the gene seat 6. In this case, the coupling coefficients W and threshold values θ are arrayed on the linear gene 5 at random. The values of the gene seat 6 are real numbers. The real number values are regarded as one element and intersection operation between genes is performed as usual. In mutation operation between next genes, an optional random number is substituted for a value as to a gene seat 6 whose mutation is determined to perform conversion to a value different from the current value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applicable to various controls such as recognition of images and voices, adaptive control of robots, associative memory, non-linear prediction, etc. Regarding the device.

[0002]

2. Description of the Related Art The function of a nerve cell (neuron), which is a basic unit of information processing in a living body, is mimicked, and further, this "nerve cell mimicking element" (nerve cell unit) is networked to perform parallel processing of information. What we aimed for was a so-called neural network. Although it is easy to perform character recognition, associative memory, motion control, etc. in a living body, there are many things that conventional Neumann computers cannot easily achieve. Attempts have been actively made to solve these problems by imitating the neural system of a living body, in particular, the functions peculiar to the living body, that is, parallel processing, self-learning, etc. by a neural network.

Here, in this kind of neural network, in order to optimize its structure, it is necessary to appropriately set parameters such as a coupling coefficient between neuron layers and a threshold value of a neuron. Therefore, as a method of updating parameters such as the coupling coefficient and the threshold value, for example, Japanese Patent Laid-Open No. 6-187473 discloses that a genetic algorithm (Genetic Algorithm) is used.

That is, the above-mentioned publication is intended to realize a genetic algorithm capable of generating a child gene that optimally inherits the features of the parents' genes and suppressing the increase in the amount of calculation as much as possible. Before the gene crossover, the genes of the parents are classified and sorted according to their similarity, so that the genes of the children that inherit the parents' genes almost equally are generated.

[0005]

However, according to the above publication, only the principle of using a genetic algorithm is shown, and even how to treat specific parameters such as a coupling coefficient and a threshold as a gene. Is not mentioned. Therefore, it is not clear how to implement the parameter updating device using the genetic algorithm in the neural network only by the technique disclosed in the above publication.

[0006]

According to the first aspect of the present invention,
Regarding a gene in which the coupling coefficient and the threshold value expressed by real numbers are randomly arranged as the value of the locus, it is configured to have a genetic algorithm calculating means for replacing the value of the locus with an arbitrary random number during mutation processing. .

According to a second aspect of the present invention, for a gene having a locus value of a coupling coefficient expressed by an N-bit code and a threshold value, an arbitrary random number is added to the locus value during mutation processing. The configuration has a genetic algorithm calculation means.

According to a third aspect of the present invention, a coupling coefficient represented by an N-bit code and a threshold value are arranged in one dimension.
Regarding the gene of the book, it was constructed to have a genetic algorithm calculating means for cutting an arbitrary part of this gene at the time of crossover processing.

According to a fourth aspect of the present invention, a coupling coefficient represented by an N-bit code and a threshold value are arranged in one dimension.
With respect to the gene of the book, it was configured to have a genetic algorithm calculation means for cutting an arbitrary part of this gene at the time of crossover processing and masking at the time of mutation processing.

[0010]

According to the first aspect of the invention, a gene is constructed by randomly arranging the coupling coefficient as a parameter and the threshold value in a real number expression, and a genetic algorithm operation is performed on a certain locus of such a gene. The means updates the parameters by performing the replacement process with an arbitrary random number. As a result, the genetic algorithm can be easily implemented to update the coupling coefficient and threshold value of the neural network, and as a result, the ability to search for the optimum parameter is also improved.

In a second aspect of the present invention, a gene is constructed by arranging a coupling coefficient as a parameter and a threshold value in an N-bit code representation, and a genetic algorithm is applied to a certain locus of such a gene. Arbitrary random number addition processing is performed by the calculation means and the parameters are updated. This gives the neural network coupling coefficient,
The genetic algorithm can be easily implemented to update the threshold parameter, and as a result, the ability to search for the optimum parameter is also improved. In particular, since the code representation is used, the consistency when the neural network is configured as a digital signal processing system is also improved.

According to the third aspect of the invention, the coupling coefficient as a parameter and the threshold value are 1 in N-bit code representation.
One gene is formed by being arranged in a dimension, and by cutting an arbitrary part of such a gene during the crossover process by the genetic algorithm computing means, it is subjected to the mutation process and the parameters are updated. . As a result, the genetic algorithm can be easily implemented to update the coupling coefficient and threshold value of the neural network, and as a result, the ability to search for the optimum parameter is also improved. In particular, since the code representation is used, the consistency when the neural network is configured as a digital signal processing system is also improved.

In the invention described in claim 4, claim 3
In addition to the crossover process according to the described invention, the masking process is performed by the genetic algorithm calculating means during the subsequent mutation process to update the parameters. Therefore,
In addition to the effect of the invention described in claim 3, the ability to search for the optimum parameter is improved.

[0014]

【Example】

<Introduction> Basically, the genetic algorithm used in the present invention is basically the following processing procedure Gene evaluation of genes (also called “individuals”) Gene evaluation, selection of termination genes if termination conditions are satisfied Crossover between genes It is done by returning to the mutation of the gene.

In the present invention, as shown in FIG. 2, the coupling coefficient and the threshold value in the neural network 1 are updated by the processing of the genetic algorithm computing means 2, and the updated coupling coefficient and the threshold value are updated in the neural network 1. Configured to load.

For example, as illustrated in FIG. 3, an input pattern I _j and an output pattern O _j are generated by a neural network 1 formed by connecting a plurality of neurons 3 hierarchically.
The case of conversion to will be described as an example. There are M patterns in total, j is an integer from 0 to (M-1),
The M number of pattern conversions are stored in the neural network 1. When the input pattern I _j is input to the neural network 1, the parameter updating device 4 including the genetic algorithm calculation means 2 sets the parameters in the following procedure (details of the above procedure) so that O _j is output as the output pattern. Will be gradually updated.

L genes (parameters) are generated using random numbers. A gene k is loaded into neural network 1. In this neural network 1, the input I
_If the output for _j is _O'j ,

[Equation 1] Is calculated. k is an integer from 0 to (L-1),
Calculate δ (k) for all genes, ie for k = 0 to (L-1). This δ (k) is the fitness of the gene k. If the fitness δ (k) is δ (k) ≤ ε, the gene k becomes the optimum parameter and the process ends. A gene has a probability L proportional to the reciprocal of fitness δ (k).
Select one. In this case, they may overlap. Perform crossover operation by pairing genes. For each locus, mutation is performed with probability R. Return to.

By executing such processing, optimum parameters of the neural network 1 can be obtained.
However, the above-described example of calculating the fitness is an example, and the present invention is not limited to this.

With regard to such processing, in the genetic algorithm of the present invention, as shown in FIG. 4, the coupling coefficient W of the neurons 3 constituting the neural network 1 is determined.
And the threshold value θ are components of gene 5. Then, one gene 5 is configured with all the coupling coefficients W in the neural network 1 and the threshold value θ as values of the locus 6.

With respect to such a gene 5, in each of the embodiments of the present invention, the gene 5 is constructed as follows, and the following genetic algorithm processing is performed to obtain the coupling coefficient W and the threshold value. Used for updating θ.

<Detailed Description> An embodiment of the invention described in claim 1 will be described. In this case, the coupling coefficient W and the threshold value θ are
They are randomly arranged on the one-dimensional gene 5. The value of locus 6 is a real number. Then, using this real value as one element, the crossover operation (crossover process) of is performed as usual. Then, in the next mutation operation (mutation processing), the locus 6 determined to cause mutation is converted into a value different from the current value by substituting an arbitrary random number.

In the mutation operation, the mutated ratio of each gene 5 (locus 6) is an empirically determined value (a value that gives the best result by repeating the experiment). If it cannot be determined empirically,
Values below 0.1 are used.

For example, in FIG. 1, real numbers a, b,
Regarding the gene 5 before mutation by the sequences of c, d, and e, it is determined that the fourth locus 6 is determined as a mutation, and the mutation operation replaces the real number d with a different random number f. Shows.

An example of application to the neural network 1 will be described here. For example, it is assumed that the parameters included in the neuron 3 of the neural network 1 as shown in FIG. 2 are the coupling coefficients; W ₁₁ , W ₁₂ , W ₂₁ , W ₂₂ thresholds; θ ₁ , θ ₂ . Gene 5 is expressed by arranging these six parameters in one dimension. The order of arranging these parameters at the time of arrangement is arbitrary, and they may be arranged at random. Therefore, a, b, c, d, e in FIG.
May be gene 5 expressed as in W ₂₁ , θ ₁ , W ₁₁ , θ ₂ , W ₁₂ , W ₂₂ ………… (a), or θ ₂ , W ₁₂ , It may be gene 5 expressed as W ₁₁ , W ₂₂ , θ ₁ , W ₂₁ (b).

For example, with specific numerical values, W ₁₁ =
0.1, W ₁₂ = 0.2, W ₂₁ = -0.3, W ₂₂ = 0.
6, θ ₁ = 0.5, θ ₂ = −0.2, when gene 5 has the sequence (a), −0.3, 0.5, 0.1, −0.2, 0. A real number of 2,0.6 is the gene 5 arranged as each locus 6.

An embodiment of the invention described in claim 2 will be described. In this case, the value of the locus 6 (coupling coefficient W and threshold θ) is represented by an N-bit code (N is an arbitrary natural number). Also, this coding may be a binary code, a Gray code, or some other code. The crossover operation is performed by using such an N-bit code as one element. And
In the next mutation operation, an arbitrary random number (code expression) is added to the current value with respect to the locus determined to cause the mutation.

For example, when the value is a binary code of N = 4 and there is a gene 5 having a code sequence such as | 1000 | 0010 | 1001 | 1111 |, as shown in FIG. The third locus 6 was determined to cause a mutation, and the code "1001" of this locus 6 was changed to "010
A random number having a code expression of "1" is added, and after mutation, a gene 5 as shown in FIG. 5B is obtained.

An embodiment of the invention described in claim 3 will be described. In this case, the coupling coefficient W and the threshold value θ are represented by an N-bit code (N is an arbitrary natural number). Then, such an N-bit code is one-dimensionally arranged as shown in FIG. 6 to form one gene 5 having an M-bit sequence. In FIG. 6, N = 4, and “1000” and “0010”.
An example is shown in which codes "1001" and "1111" are arranged in order. In such a crossover operation for one gene 5, an arbitrary position in the M bit string is cut. In FIG. 6, reference numeral 7 indicates a cutting point. The cutting point 7 is randomly selected. Subsequent processing is the same as that of a normal genetic algorithm, and a mutation operation and the like are performed.

An embodiment of the invention described in claim 4 will be described. This embodiment is different from the embodiment of the invention described in claim 3 in that before the mutation, a M-bit mask 8 as shown in FIG. The mask processing is performed on the gene 5. “0” in mask 8
For the part marked with, the corresponding bit of gene 5 is inverted, and the part with "1" is left unchanged. By such masking, the gene 5 after mutation is shown in FIG.
It is converted as shown in (c).

[0030]

INDUSTRIAL APPLICABILITY As described above, the present invention clarifies how a gene is constituted by a coupling coefficient as a parameter and a threshold, and clarifies mutation treatment or crossover treatment for such a gene. Therefore, the genetic algorithm can be easily implemented for updating the parameters of the neural network, and as a result, the ability to search for the optimum parameter can be improved.
According to the inventions described in (4) to (4), since the gene is configured by using the parameter as the code expression, it is possible to improve the consistency when the neural network is configured as the digital signal processing system.

[Brief description of drawings]

FIG. 1 is a schematic diagram showing an embodiment of the invention described in claim 1.

FIG. 2 is a block diagram showing a schematic configuration of the entire system.

FIG. 3 is a block diagram showing a more specific version of FIG.

FIG. 4 is a schematic diagram showing gene constitution.

FIG. 5 is a schematic view showing an embodiment of the invention described in claim 2.

FIG. 6 is a schematic view showing an embodiment of the invention described in claim 3.

FIG. 7 is a schematic diagram showing an embodiment of the invention described in claim 4.

[Explanation of symbols]

 2 Genetic algorithm operation means 5 genes 6 loci 8 masks

Claims (4)

[Claims] 1. A genetic algorithm arithmetic means for replacing a gene of a gene in which a coupling coefficient expressed by a real number and a threshold value are randomly arranged as a gene locus value and replacing the gene locus value with an arbitrary random number during mutation processing. A parameter updating device for a neural network, comprising: 2. A genetic algorithm calculating means for adding an arbitrary random number to the value of the locus at the time of mutation processing for a gene having a locus value with a coupling coefficient expressed by an N-bit code and a threshold value. A neural network parameter updating device having. 3. A single gene in which a coupling coefficient expressed by an N-bit code and a threshold value are arranged one-dimensionally,
A parameter updating device for a neural network, which has a genetic algorithm calculating means for cutting an arbitrary part of this gene at the time of crossover processing. 4. One gene in which a coupling coefficient expressed by an N-bit code and a threshold value are arranged one-dimensionally,
A parameter updating device for a neural network, which has a genetic algorithm calculating means for cutting an arbitrary part of this gene during a crossover process and performing a mask process during a mutation process.