JP2002245434A - Device and program for searching optimal solution by evolutionary method and controller of control object by evolutionary method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a controller of a control object by an evolutionary method suitable to obtain an effective solution with a small number of the alternation of generations. SOLUTION: The fuel consumption characteristic and response characteristic of an engine 10 are optimized for a user by a GA(Generic Algorithm). The GA virtually generates a population composed of the sets of a plurality of individuals, and configures individual information regarded as the genetic information of the individual for each individual. Here, a control coefficient for constructing a control module is allocated to each individual information. In the same generation, the population is made to evolve by performing a genetic operation that performs an information operation simulating a genetic operation on the individual information, and an individual selection operation that performs individual existence or selection on the basis of the evaluation value of an individual by at least once each to progress a generation by several times. In the process of generation progress, the degree of response is predicted on the basis of individual information for each individual, and the genetic operation is applied to the individual information until the predicted degree of response satisfies a prescribed condition.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus and a program based on an evolutionary method constructed by imitating the process of natural evolution, and more particularly to an evolutionary method suitable for obtaining an effective solution with a small number of generation alternations. The present invention relates to an optimal solution search device based on an ERP, a control device for a controlled object by an evolutionary method, and an optimal solution search program by an Evolutionary method.

[0002]

2. Description of the Related Art Conventionally, when controlling the characteristics of a product such as a vehicle or a home appliance, the characteristics of the product to be controlled are assumed at a development / design stage by assuming a user who will use the product. Taking into account the preferences and usage conditions of the virtual user, the virtual user is determined to adapt to as wide a range of users as possible. However, the users who use the above products have their own unique personalities, and their preferences vary widely, so as described above, the preferences of the users who are likely to use the products, etc. Even if products are developed and designed on the assumption, it is almost impossible to provide characteristics that are satisfied by all users. In order to solve this problem, the user's preference and use situation are estimated after purchase using a neural network or a genetic algorithm (hereinafter simply referred to as GA), and the control characteristics are adjusted to characteristics that can satisfy the user. A changing control method is being attempted. This includes, for example,
The one disclosed in Japanese Patent Publication No. 03 is known.

[0003] GA is a stochastic search that performs an efficient search by imitating the process of natural evolution and evolving an individual representing a search point in a search space toward an optimal solution based on the principle of survival of the fittest. Algorithm. More specifically, an individual group consisting of a set of a plurality of individuals is virtually generated, and individual information is configured for each individual in terms of genetic information of the individual. Then, in the same generation, a genetic operation in which an information operation imitating a genetic operation is performed on individual information and an individual selection operation in which an individual survives or is eliminated based on an evaluation value of an individual are performed at least once, respectively. To progress. Although it varies somewhat depending on the application target, in order to obtain a somewhat effective solution, it is generally necessary to repeat several hundred to several thousand generation changes.

[0004]

However, for example, when GA is applied to optimize the control characteristics of a vehicle engine for a user, and the user gives the ride comfort as an evaluation value, In order to perform one generation change, the evaluation by the user must be performed without fail.
The user has to give hundreds to thousands of evaluations to obtain satisfactory characteristics, which takes a lot of time and is very troublesome.

Therefore, it is desired that an effective solution can be obtained with as few generational alternations as possible. Of course, this can be said to be not only a case where the control characteristics of the vehicle engine are optimized, but also a general problem that the GA has. Therefore,
The present invention has been made in view of such unresolved problems of the conventional technology, an optimal solution search device by an evolutionary method suitable for obtaining an effective solution with a small number of generation alternations, An object of the present invention is to provide a control device for a controlled object by an evolutionary method and an optimal solution search program by an evolutionary method.

[0006]

In order to achieve the above object, an optimal solution search device according to the present invention according to the first aspect of the present invention virtually generates an individual group consisting of a set of a plurality of individuals. At the same time, individual information is configured for each individual based on the genetic information of the individual, and in the same generation, an information operation simulating the genetic operation is performed on the individual information by a genetic operation and an evaluation function. An apparatus for searching for an optimal solution of the evaluation function by performing an individual selection operation of performing survival or selection of the individual based on the evaluation value of the individual at least once, and causing the generation to proceed, A value is predicted, and the genetic operation is performed based on the predicted value.

With such a configuration, the generation proceeds by performing the genetic operation and the individual selecting operation at least once each in the same generation. In the process of generation progress, an evaluation value is predicted, and a genetic operation is performed based on the predicted value. For example, as one form,
In the same generation, a prediction and a genetic operation based on the prediction are performed one or more times, and thereafter, an individual selection operation is performed, whereby the generation proceeds.

[0008] Here, the prediction and the genetic operation based on the prediction may be performed separately from the genetic operation and the individual selection operation performed at least once in the same generation,
The genetic operation may be performed at least once in the same generation. When the process is performed separately, the process need not always be performed in the same generation. For example, the process may be performed every predetermined generation. Hereinafter, the same applies to the control device of the control target by the evolutionary method according to claim 6 and the optimal solution search program by the evolutionary method according to claim 15.

The estimation of the evaluation value may be performed based on the individual information, or may be performed based on information unrelated thereto. Hereinafter, the same applies to the optimal solution search program by the evolutionary method according to claim 15. Further, the invention of claim 1 can be realized as a GA, and in addition, a GP (Genetic Progress)
And ES (evolutional Stratage). Hereinafter, the same applies to the control device of the control target by the evolutionary method according to claim 6 and the optimal solution search program by the evolutionary method according to claim 15.

Further, according to a second aspect of the present invention, there is provided an optimal solution searching apparatus based on the evolutionary method according to the first aspect, wherein the prediction value satisfies a predetermined condition until the predicted value satisfies a predetermined condition. The genetic operation and the prediction are repeatedly performed. With such a configuration, the genetic operation and the prediction are repeatedly performed until the predicted value satisfies the predetermined condition.

Further, according to a third aspect of the present invention, there is provided an optimal solution searching apparatus using an evolutionary method according to any one of the first and second aspects. A prediction unit configured to predict the second evaluation value based on the individual information, the prediction unit comprising a first evaluation value and a second evaluation value; The operation and the prediction are repeatedly performed.

With such a configuration, the second evaluation value is predicted by the prediction means based on the individual information, and the genetic operation and the prediction are repeatedly performed until the predicted value satisfies a predetermined condition. Furthermore, the optimal solution search device using the evolutionary method according to claim 4 of the present invention is the third embodiment.
In the optimal solution search device according to the described evolutionary method, the genetic operation is performed on information affecting the second evaluation value among the individual information based on the predicted value.

With such a configuration, a genetic operation is performed on information affecting the second evaluation value in the individual information based on the predicted value. Furthermore, the optimal solution search device by the evolutionary method according to claim 5 according to the present invention is the optimal solution search device by the evolutionary method according to any one of claims 3 and 4, wherein the predicting means is configured to perform the learning by learning. The evaluation value is predicted, and learning is performed based on the difference between the predicted value and the evaluation value.

With such a configuration, the estimation means predicts the evaluation value by learning, and the learning is performed based on the difference between the predicted value and the evaluation value. On the other hand, in order to achieve the above object, the control device of the control object by the evolutionary method according to claim 6 according to the present invention virtually generates an individual group consisting of a set of a plurality of individuals, For each individual, the individual information is configured based on the genetic information of the individual, and the individual information is assigned a control coefficient that influences the control characteristics of a control system that controls the characteristics of the control target, and further imitates genetic manipulation. Individual information operation means for performing information operations on the individual information; evaluation value calculation means for calculating an evaluation value of the individual; and survival or selection of the individual based on the evaluation value calculated by the evaluation value calculation means. Individual selecting means for performing the genetic operation by the individual information operating means and the individual selecting operation by the individual selecting means at least once in the same generation to advance the generation. ,
An apparatus for optimizing the control characteristics of the control system by using the individual information of the individual that has evolved as the control coefficient, and a prediction unit that predicts the evaluation value based on the individual information. The genetic operation is performed based on the predicted value predicted by the prediction means.

With such a configuration, the individual information operating means performs an information operation imitating the genetic operation on the individual information, the evaluation value is calculated by the evaluation value calculating means, and the individual value selecting means is used by the individual selecting means. Survival or selection of the individual is performed based on the calculated evaluation value. The generation proceeds by performing the genetic operation by the individual information operation means and the individual selection operation by the individual selection means at least once in the same generation. In the process of generation progression, an evaluation value is predicted by the prediction means, and a genetic operation is performed based on the predicted value. As one form, for example, in the same generation, a prediction and a genetic operation based on the prediction are performed one or more times, and then an individual selection operation is performed, so that the generation proceeds. When the individual finally evolves in this way, the individual information of the individual is used as a control coefficient.

Further, according to a seventh aspect of the present invention, there is provided a control apparatus for controlling an object by an evolutionary method, wherein the evaluation value is a first evaluation value. And the second evaluation value, wherein the prediction means is configured to predict the second evaluation value based on the individual information, and until the prediction value predicted by the prediction means satisfies a predetermined condition. The genetic operation and the prediction are repeatedly performed.

With such a configuration, the second evaluation value is predicted by the prediction means based on the personal information, and the genetic operation and the prediction are repeatedly performed until the predicted value satisfies a predetermined condition. Furthermore, the control device of the controlled object by the evolutionary method according to claim 8 according to the present invention is the control device of the controlled object by the evolutionary method according to claim 7, wherein the control unit controls the individual information based on the predicted value. The genetic operation is performed on information affecting the second evaluation value.

With such a configuration, a genetic operation is performed on information affecting the second evaluation value in the individual information based on the predicted value. Further, according to a ninth aspect of the present invention, there is provided the control device for an object to be controlled by an evolutionary method, wherein the control object is an engine. Wherein the first evaluation value is a fuel efficiency of the engine, and the second evaluation value is a responsiveness determined by a rotation rate change rate and a throttle opening change rate of the engine.

With such a configuration, the predicting means calculates the response based on the individual information, and the evaluation value calculating means calculates the fuel efficiency and the response. Further, according to a tenth aspect of the present invention, in the control apparatus for controlling a controlled object by the evolutionary method according to the ninth aspect, in the control apparatus for a controlled object by the evolutionary method according to the ninth aspect, the predicting means includes: The second evaluation value is predicted based on a fuel injection amount of the engine and a transient correction amount for correcting the fuel injection amount in a transient state of the engine, which is obtained when the engine is operated using Has become.

With such a configuration, the response is set as the second evaluation value by the prediction means based on the fuel injection amount and the transient correction amount obtained when the engine is operated using the individual information as the control coefficient. is expected. Furthermore, the control device of the control target by the evolutionary method according to claim 11 according to the present invention is the control device of the control target by the evolutionary method according to any one of claims 6 to 10,
The control target is an engine, and the individual information includes
As the control coefficient, a fuel injection amount of the engine, a transient correction amount for correcting the fuel injection amount in a transient state of the engine, a correction value of the fuel injection amount, or a correction value of the transient correction amount is assigned. I have.

With this configuration, as the population evolves in a direction to improve the evaluation value, it is expected that a higher evaluation value can be obtained. The value or the correction value of the transient correction amount is determined. Furthermore, the control device of the controlled object by the evolutionary method according to claim 12 of the present invention is the control device of the controlled object by the evolutionary method according to any one of claims 6 to 10, wherein the controlled object is an engine. The fuel injection amount of the engine, a transient correction amount for correcting the fuel injection amount in a transient state of the engine, a correction value of the fuel injection amount, or a correction value of the transient correction amount is generated by a neural network. The individual information is assigned a synapse coupling coefficient in the neural network as the control coefficient.

With this configuration, the fuel injection amount, the transient correction amount, the correction value of the fuel injection amount, or the correction value of the transient correction amount are generated by the neural network. As the group evolves,
A synapse coupling coefficient in the neural network that can be expected to obtain a high evaluation value is determined. Further, according to a thirteenth aspect of the present invention, there is provided a control device for a controlled object by an evolutionary method, wherein the controlled object is an electric control device, A motor, wherein the first evaluation value is:
Power consumption of the electric motor, the second evaluation value is:
It is a rotation change rate of the electric motor.

With such a configuration, the prediction means calculates the rotation change rate based on the individual information, and the evaluation value calculation means calculates the power consumption and the rotation change rate. Further, the control device of the controlled object by the evolutionary method according to claim 14 of the present invention is the control device of the controlled object by the evolutionary method according to any one of claims 6 to 13, wherein the predicting means includes a learning device. And the learning is performed based on the difference between the predicted value and the evaluation value.

With such a configuration, the prediction means predicts the evaluation value by learning, and the learning is performed based on the difference between the predicted value and the evaluation value. On the other hand, in order to achieve the above object, an optimal solution search program by an evolutionary method according to claim 15 of the present invention virtually generates an individual group consisting of a set of a plurality of individuals, To compose individual information based on the genetic information of the individual,
In the same generation, a genetic operation that performs an information operation imitating a genetic operation on the individual information, and an individual selection operation that performs survival or selection of the individual based on an evaluation value of the individual obtained by an evaluation function, respectively. A computer-executable program for searching for an optimal solution of the evaluation function by performing the generation at least once, and predicting the evaluation value, and performing the genetic operation based on the predicted value. It has become.

With such a configuration, when the computer executes the program in accordance with the program, an operation equivalent to that of the optimal solution search device according to the first aspect of the invention is obtained.

[0026]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 to FIG. 12 are diagrams showing an embodiment of an optimum solution search device by an evolutionary method, a control device of a control target by an evolutionary method, and an optimum solution search program by an evolutionary method according to the present invention.

First, the basic concept of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing the basic concept of the present invention. As shown in FIG. 1, the basic configuration of the present invention includes an optimization target 1, an evaluation unit 2 that calculates an evaluation value of the optimization target 1 based on an operation result of the optimization target 1, A prediction unit 3 for predicting an evaluation value, an evaluation value calculated by the evaluation unit 2,
An evolution mechanism 4 for determining and outputting the operation amount of the optimization target 1 by an evolutionary optimization algorithm based on the prediction value predicted by the prediction unit 3 and the operation result of the optimization target 1.

The predicting section 3 has a learning mechanism, and makes the learning mechanism learn based on the evaluation value calculated by the evaluation section 2 and the operation amount calculated by the evolution mechanism 4, and the learning mechanism evaluates the evaluation value of the optimization target 1. Is to predict. Evolution mechanism 4
Is based on the evaluation value calculated by the evaluation unit 2, the prediction value predicted by the prediction unit 3, and the operation result of the optimization target 1, and the operation amount that optimizes the operation characteristic of the optimization target 1 is determined by GA. The determined operation amount is output to the optimization target 1. In the evolution simulation by GA, the genetic operation on the individual information and the prediction by the prediction unit 3 are repeatedly performed until the predicted value from the prediction unit 3 satisfies a predetermined condition in the same generation.

Next, the basic configuration of the present invention will be described with reference to FIG. FIG. 2 is a block diagram showing a basic configuration of the present invention. The basic configuration of the present invention is as shown in FIG.
The control device 50 includes a control target 50 and a control device 60 that controls a control amount of the control target 50 based on a control result of the control target 50.

The control device 60 includes a reflection layer 500, a learning layer 6
00 and an evolutionary adaptation layer 700,
A control result is input from the control target 50, a basic control amount is determined by the reflection layer 500 based on the input control result, and a correction rate for the basic control amount is determined by the learning layer 600 and the evolution adaptive layer 700. The final control amount is determined from the amount and the correction rate. Hereinafter, the reflection layer 500 and the learning layer 6
00 and the configuration of the evolution adaptive layer 700 will be described in detail.

The reflection layer 500 includes a basic control unit 510 for defining a relationship between a basic control amount and a control result in the form of a mathematical expression, a map, a neural network, a fuzzy rule, a subsumption architecture, and the like.
The control unit 10 receives a control result from the control target 50, determines a basic control amount based on the input control result, and outputs the basic control amount. The subsumption architecture is known as behavioral artificial intelligence that performs parallel processing.

The evolution adaptation layer 700 includes an interface unit 710 for input / output with the user, an evolution adaptation unit 720 for optimizing the control characteristics of the control target 50 by performing an evolution simulation using a GA, and an individual evaluation in the GA. The evaluation unit 730 includes an evaluation unit 730 that calculates a value, and a prediction unit 740 that estimates an evaluation value of an individual in GA. The evolution adaptation unit 720 has at least one control module that outputs an evolution correction rate for correcting the basic control amount from the reflective layer 500 to a value according to the user's desire based on the control result. To optimize the control module. In the GA, an individual group consisting of a set of a plurality of individuals is virtually generated, and individual information is configured for each individual in terms of genetic information of the individual. Here, a control coefficient for constructing a control module is assigned to each individual information. Then, in the same generation, a genetic operation in which an information operation imitating a genetic operation is performed on individual information and an individual selection operation in which an individual survives or is eliminated based on an evaluation value of an individual are performed at least once, respectively. Is advanced a predetermined number of times to evolve the population. When the predetermined number of generations has been changed, an individual having the highest evaluation value is extracted from the individual group, and a control module is constructed using individual information of the extracted individual as a control coefficient. Hereinafter, the control module constructed using the individual information of the individual with the highest evaluation value is referred to as the “optimal control module”.
That. Note that the control module is a unit that performs a certain control of a control system.

After constructing the optimal control module, the evolution adaptation unit 720 fixes the control module of the evolution adaptation unit 720 to the optimal control module, and adjusts the basic control amount from the reflection layer 500 by the evolution correction rate. While performing the control, the learning layer 600 is made to learn information on the optimal control module. After the learning layer 600 learns the information on the optimal control module, the output is returned to “1”, and thereafter, the operation is performed according to the user's instruction. That is, the control by the control module of the evolution adaptation unit 720 is performed only during the evolution simulation and during the learning.

The learning layer 600 includes a learning section 6 having two neural networks that can be switched between learning and execution.
The learning unit 610 is configured to control the relationship between the input and the output of the optimal control module from the evolution adaptive layer 700 while performing the control using one neural network (for execution) and the other neural network (for learning). To learn. When the learning by the learning neural network is completed, the neural network performing the control and the neural network after the learning are switched, and the control by the control module obtained from the learning result in the neural network after the learning is started. The running neural network begins to function for learning. Note that the neural network in the learning layer 600 is
In the initial state, it is set to output "1",
Therefore, in the initial state, control by the reflective layer 500 and the evolution adaptive layer 700 is performed.

The execution neural network receives a control result from the control target 50 and outputs a learning correction rate for correcting the basic control amount from the reflection layer 500 based on the input control result. . This configuration is the same for the learning neural network. Then, the control device 60 adds the learning correction rate from the learning layer 600 and the evolution correction rate from the evolution adaptation layer 700, and multiplies the basic control amount from the reflection layer 500 by the addition result to obtain the control amount. calculate. This control amount is set as the control target 50
Output to

Hereinafter, more specific embodiments of the present invention will be described. In the present embodiment, an optimal solution search device by an evolutionary method, a control device of an object to be controlled by an evolutionary method, and an optimal solution search program by an evolutionary method according to the present invention,
As shown in FIG. 3, the present invention is applied to a case where the fuel economy characteristics and response characteristics of the engine 10 are optimized for the user by GA.

First, the configuration of an engine control system to which the present invention is applied will be described with reference to FIG. FIG. 3 is a block diagram showing a configuration of an engine control system to which the present invention is applied. As shown in FIG. 3, the engine control system detects the operating state of the engine 10 and various sensors 20 that output various kinds of information on the operating state of the engine 10 (hereinafter collectively referred to as external information).
And a control device 30 that controls the fuel injection amount of the engine 10 based on the external information from the various sensors 20.

The various sensors 20 detect the operating state of the engine 10 and the running state of the vehicle, and based on the detection results, the rotation speed of the engine 10, the throttle opening, the rate of change of the throttle opening, the distance pulse, and the fuel. The injection amount is output as external information. Control device 30
Are the reflection layer 100, the learning layer 200, and the evolution adaptive layer 30
0, and inputs external information from various sensors 20. The reflection layer 100 is formed based on the input external information.
Determines the basic injection amount of the fuel, the learning layer 200 and the evolution adaptation layer 300 determine the correction amount for the basic injection amount, and determines the final fuel injection amount from the basic injection amount and the correction amount. Hereinafter, the configurations of the reflection layer 100, the learning layer 200, and the evolution adaptive layer 300 will be described in detail.

The reflection layer 100 includes a basic control unit 110 for defining the relationship between the basic injection amount and the transient correction rate and the external information in the form of a mathematical expression, a map, a neural network, a fuzzy rule, a subsumption architecture, and the like. The control unit 110 inputs the external world information from the various sensors 20 and determines and outputs the basic injection amount and the transient correction rate based on the input external world information.

The evolution adaptation layer 300 includes an interface unit 310 for performing input / output with the user, an evolution adaptation unit 320 for optimizing the control characteristics of the engine 10 by performing an evolution simulation using a GA, and an individual evaluation value for the GA. And an estimating unit 340 that estimates an individual's evaluation value in GA. The evolution adaptation unit 320 corrects the basic injection amount and the transient correction rate from the reflective layer 100 to values according to the user's desire based on the external world information (hereinafter, the basic injection amount of the correction rate). , And at least one control module that outputs a transient correction rate is referred to as an evolutionary correction rate, and the GA is configured to optimize the control module by GA. ing.

After constructing the optimal control module, the evolution adaptation unit 320 fixes the control module of the evolution adaptation unit 320 to the optimal control module, and adjusts the basic injection amount from the reflection layer 100 based on the evolution correction rate. While the control and the control based on the evolutionary transient correction rate for correcting the transient correction rate from the reflection layer 100 are performed, the learning layer 200 is made to learn information on the optimal control module. After the learning layer 200 learns the information on the optimal control module, the output is returned to “1”, and thereafter, the operation is performed according to the user's instruction. That is, the control by the control module of the evolution adaptation unit 320 is performed only during the evolution simulation and during the learning.

The learning layer 200 includes a learning unit 2 having two neural networks that can be switched between learning and execution.
10, the learning unit 210 controls the relationship between the input and the output of the optimal control module from the evolution adaptive layer 300 in the other neural network (for learning) while performing the control in one neural network (for learning). To learn. When the learning by the learning neural network is completed, the neural network performing the control and the neural network after the learning are switched, and the control by the control module obtained from the learning result in the neural network after the learning is started. The running neural network begins to function for learning. The neural network in the learning layer 200 is
In the initial state, it is set to output "1",
Therefore, in the initial state, control by the reflective layer 100 and the evolution adaptive layer 300 is performed.

Although not shown, the execution neural network further includes two neural networks. On the other hand, the neural network inputs a throttle opening and an engine speed as various kinds of external information from various sensors 20 and corrects a basic injection amount from the reflective layer 100 based on the input information. The other neural network inputs the rate of change of the throttle opening and the engine speed from the various sensors 20 as external information, and based on the input information. A correction rate for correcting the transient correction rate from the reflective layer 100 (hereinafter, this correction rate is referred to as a learning transient correction rate) is output. This configuration is the same for the learning neural network.

Then, the control device 30 adds the learning correction rate from the learning layer 200 and the evolution correction rate from the evolution adaptive layer 300, and multiplies the basic injection amount from the reflection layer 100 by the addition result. Is the first multiplication result, the learning transient correction rate from the learning layer 200 and the evolutionary transient correction rate from the evolution adaptive layer 300 are added, and the transient correction rate from the reflection layer 100 is multiplied by the addition result. This is used as the second multiplication result, and the fuel injection amount is calculated by multiplying the first multiplication result by the second multiplication result. This fuel injection amount is output to the engine 10.

Next, the configuration of the evolution adaptive layer 300 will be described in detail with reference to FIG. FIG. 4 shows the evolution adaptation layer 300.
FIG. 3 is a block diagram showing the configuration of FIG. The evolution adaptation layer 300
As shown in FIG. 4, the configuration includes an interface unit 310, an evolution adaptation unit 320, an evaluation unit 330, and a prediction unit 340. The interface unit 310 includes a display unit 312 for displaying an evaluation value of the individual during the evolution simulation by the GA, and an input unit 31 for inputting the evaluation by the user.
During the evolutionary simulation by GA, the evaluation value of each individual (fuel consumption and response level, which will be described in detail later) is displayed on the display unit 312 for each generation, and the user can ride the vehicle. The evaluation value of each individual is input to the input unit 314 based on the experience of the vehicle such as comfort.

The evaluation unit 330 calculates a fuel consumption of the engine 10 based on the fuel injection amount and the distance pulse, and a response calculation calculates a response based on the throttle opening and the engine speed, which are external information. And a degree calculation unit 334. Fuel efficiency calculation unit 33
2 is to input the fuel injection amount and the distance pulse as external information, calculate the fuel consumption by summing the injection amount at the input interval of the distance pulse input every time the vehicle travels a predetermined distance, and calculate the calculated fuel efficiency as the individual Is output to the evolution adaptation unit 320 as the first evaluation value of The response degree calculation unit 334 inputs the throttle opening and the engine speed as external information, calculates the throttle opening change rate and the engine speed change rate, and calculates the engine speed change rate as the throttle opening change. The responsivity is calculated by dividing by the ratio, and the calculated responsivity is referred to as GA
Is output to the evolution adaptation unit 320 and the prediction unit 340 as the second evaluation value of the individual at.

The prediction unit 340 is a CMAC (Cerebellar M
odel Arithmetic Computer) 342 and the reflective layer 10
The learning layer 200 and the evolution adaptive layer 3 are added to the basic injection amount from 0.
A value obtained by multiplying the correction rate from 00 (hereinafter, referred to as a corrected basic injection amount), and a value obtained by multiplying the transient correction rate from the reflective layer 100 by the correction rate from the learning layer 200 and the evolution adaptive layer 300 (hereinafter, referred to as the correction rate). , The corrected transient correction rate), the second evaluation value among the individual evaluation values in the GA is predicted. The responsivity, which is the second evaluation value, is calculated by multiplying the corrected basic injection amount by the corrected transient correction rate and adding the output of the CMAC 342 to the multiplication result. Further, the CMAC 342 learns based on an error between the measured value and the predicted value of the response level.

Specifically, the predicted value of the responsivity is obtained by calculating the initial value of the responsivity and CMAC3 by the following equation (1).
It can be calculated by adding the outputs of the C.I. In the following equation (1), Exp_resp is a predicted value of the responsivity, Ini_resp is a predicted initial value of the responsivity,
CMAC_out is the output of CMAC 342.

[0049]

(Equation 1)

The initial value of the predicted response is based on the transient fuel injection amount, and the transient fuel injection amount is obtained by multiplying the corrected basic injection amount by the corrected transient correction rate. The following equation (2) can be used to calculate by multiplying the corrected basic injection amount by the corrected transient correction rate. In the following equation (2), Statistic_f is a statistical model value of the basic injection amount, and Statistic_a is a statistical model value of the transient correction rate.

[0051]

(Equation 2)

The statistical model value of the basic injection amount and the statistical model value of the transient correction rate are statistically calculated with respect to the N points included in the hatched area in FIG. 5 used in the reference individual for maintaining the response. It is a value calculated using a statistical model. FIG. 5 is a diagram showing an area used in the reference individual. The shaded area is set as a set U (i), and the points included therein are set as i ′ (∈U (i)). The corrected basic injection amount at each point of U (i) is f (i '), the corrected transient correction rate is a (i'), and the value of the same point in the reference individual for maintaining the response is Assuming that f _b (i ′) and a _b (i ′), the statistical model value of the basic injection amount and the statistical model value of the transient correction rate are F (i) satisfying the following equations (3) and (4), respectively. '), A (i'), and finally the following equations (5) and (6)
Can be calculated by

[0053]

(Equation 3)

[0054]

(Equation 4)

[0055]

(Equation 5)

[0056]

(Equation 6)

The evolution adaptation unit 320 is provided for the control module 32
2, and the control module 322 further includes two neural networks. One neural network 322a inputs a throttle opening and an engine speed as external information from various sensors 20 and outputs an evolutionary correction rate based on the input information.
Reference numeral 22b inputs the change rate of the throttle opening and the engine speed as external information from various sensors 20, and outputs an evolutionary transient correction rate based on the input information.

The synapse coupling coefficients in the neural networks 322a and 322b are assigned to the individual information of the individual in the GA. Specifically, they are assigned as shown in FIG. FIG. 6 is a diagram showing a configuration of the neural networks 322a and 322b and a data structure of individual information. Neural network 322a
Is an input layer f _i1 for inputting the throttle opening, an input layer f _i2 for inputting the engine speed, intermediate layers f _h1 and f _h2 for inputting the outputs from the input layers f _i1 and f _i2 , and an intermediate layer f _h1 and f _h2. f _h1 , f
_It comprises five perceptrons with an output layer f _o1 that inputs the output of _h2 and outputs the evolution correction rate. And
The input layer f _i1 and the intermediate layer f _h1 are connected by the synapse of the coupling coefficient k _f1 , and the input layer f _i2 and the intermediate layer f _h1 are connected by the synapse of the coupling coefficient k _f2 , and the intermediate layer f _h1 and the output layer f _o1 are connected by the coupling coefficient. By the synapse of k _f3 , the input layer f _i1 and the intermediate layer f _h2 are connected by the synapse of the coupling coefficient k _f4 , and the input layer f _i2 and the intermediate layer f _h2 are output by the synapse of the coupling coefficient k _f5 and the output of the intermediate layer f _h2. Layer f _o2
Are connected by synapses having a coupling coefficient k _f6 .

The neural network 322b includes an input layer a _i1 for inputting the rate of change of the throttle opening, an input layer a _i2 for inputting the engine speed, and an intermediate layer for inputting the outputs from the input layers a _i1 and a _i2. a _h1 , a _h2 and the middle layer a _h1 , a
Output layer a that _receives the output of _h2 and outputs the evolutionary transient correction rate
_o1 and five perceptrons. By synaptic input layer a _i1 and coupling the intermediate layer a _h1 coefficient k _a1, the synaptic input layer a _i2 and the intermediate layer a _h1 is the coupling coefficient k _a2, the intermediate layer a _h1 and the output layer a _o1 is Coupling coefficient k
Due to the synapse of _a3 , the input layer a _i1 and the intermediate layer a _h2 are formed by the synapse of the coupling coefficient k _a4 , and the input layer a _i2 and the intermediate layer a _h2 are formed by the synapse of the coupling coefficient _{ka 5} , and the intermediate layer a _h2 and the output layer are formed. a _o2 is connected to each other by a synapse having a connection coefficient k _a6 .

The individual information of the individual in the GA is
The synapse coupling coefficients k _{f1 to} k _f6 are sequentially assigned to the upper side, and the synapse coupling coefficients k _{a1 to} k _a6 are successively assigned to the lower side. For example, if one coupling coefficient is composed of 8-bit data, the individual information becomes 96-bit data as a whole. The initial individual information generated at the start of the evolution simulation is determined by a random number for each individual. At this time, it is preferable to limit the random number generation range to a predetermined range in order to guarantee a certain degree of response. That is, random numbers are not generated in a range where the response level is clearly deteriorated.

Next, the processing executed by the evolution adaptation unit 320 will be described in detail with reference to FIG. FIG. 7 is a flowchart illustrating a process executed by the evolution adaptation unit 320. The GA assigns random initial values to each individual and arranges them in a search space, applies genetic operations called crossover and mutation for each generation, and performs growth and selection of individuals according to the evaluation value of the individual Thereby, a set of individuals of the next generation is obtained. An object is to asymptotically approach an optimal solution by repeating such generation alternation. Hereinafter, crossover, mutation, and selection, which are genetic operations, will be described.

[0062] Crossover is an operation of generating at least one descendant individual by replacing at least two individuals as parents and replacing part of the individual information of the parent individual. By combining the good part of the individual information of a certain individual with the good part of the individual information of another individual, it is expected that an individual with a higher evaluation value will be obtained. For example, in a case where two individuals as descendants are generated using two individuals as parents, the individual information of one parent individual is “000110”, and the individual information of the other parent individual is “110111”. By crossing at the position of “11,
The individual having the individual information of “0110” is obtained as an individual to be a descendant.

Mutation is an operation of changing a specific part of individual information of an individual with a predetermined probability, and increases diversity within an individual group. Specifically, this is an operation of inverting a specific bit of the individual information. For example, by setting the individual information of a certain individual to “000111” and causing a mutation at the third position, the individual information of “001111” is obtained. Obtain an individual with

The selection is an operation for leaving a better individual from the population to the next generation according to the evaluation value of the individual.
In a selection method called roulette selection, each individual is selected with a probability proportional to the evaluation value. For example, in a certain generation, “000000”, “111011”, “110111”, “01011”
The evaluation value of the individual having the individual information of “1” is “8”,
It is assumed that they are "4", "2", and "2". The probability that each individual will be selected is “8/16”, “4/16”,
"2/16" and "2/16". Therefore, on average, in the next generation, the number of individuals having the individual information of “000000” increases to two, the number of individuals having the individual information of “111011” remains one, and the individual having the individual information of “110111” remains. Alternatively, an individual group having any of the individuals having the individual information “010111” is obtained. However, in this embodiment, the individual is selected by the user.

Based on the above, the evolution adaptation unit 320
The processing executed by will be described. The process shown in the flowchart of FIG. 7 is executed, for example, by reading a program stored in the ROM in advance and executing the program according to the read program. First, the process proceeds to step S100 to determine whether or not an evolution start instruction, which is an instruction to start an evolution simulation, has been input from the input unit 314. If it is determined that the evolution start instruction has been input (Yes),
The process proceeds to step S102, but if it is determined that this is not the case (No), step S1 is performed until an evolution start instruction is input.
Wait at 00.

In step S102, a predetermined number (for example,
An individual group consisting of a set of (9) individuals is virtually generated, and individual information is configured for each individual. Here, the individual information includes the neural networks 322a and 322a.
The synapse coupling coefficient in 2b is assigned, and individual information of each individual is determined by a random number. At this time, by generating one individual in which all values of the individual information are “0”,
In the process of evolution, the response performance before evolution can be maintained. The individual information of each individual is RA
It is stored and managed on a storage device such as M.

Then, the process proceeds to step S104 to read the individual information of the first individual in the individual group, and proceeds to step S106 to determine the connection state of the neural networks 322a and 322b based on the read individual information. Then, the control module 322 is constructed, and control of the engine 10 is started by the constructed control module 322. At this time, the output from the evolution adaptation layer 300 is obtained by inputting the throttle opening, the change rate of the throttle opening and the engine speed to the neural networks 322a and 322b, and further linearly converting the output by the following equation (7). It is calculated by: The input information of the throttle opening, the change rate of the throttle opening, and the engine speed is normalized. In the following equation (7),
Y is the evolutionary correction rate or evolutionary transient correction rate, x is the output of neural networks 322a, 322b,
G is a predetermined gain.

[0068]

(Equation 7)

As described above, the neural network 32
By linearly transforming and using the output x of 2a and 322b, the value of the evolutionary correction rate or the evolutionary transient correction rate output from the evolutionary adaptation layer 300 does not become extremely large, and the evolutionary simulation proceeds little by little as a whole. Therefore, the behavior of the engine 10 does not extremely fluctuate due to evaluation or evolution simulation.

Next, the process proceeds to step S108 to acquire the predicted value of the response degree from the prediction unit 340, and proceeds to step S110 to determine whether the obtained predicted value satisfies a predetermined condition. Whether or not the predetermined condition is satisfied is calculated by calculating the error of the predicted value of the response degree with respect to the reference value by the following equation (8), and determining whether or not the error is below the allowable range. In the following equation (8), Error is the error of the predicted value with respect to the reference value, Exp_resp is the predicted value of the responsivity, and Base_r
esp is a reference value.

[0071]

(Equation 8)

As a result of the determination, if it is determined that the predicted value of the response degree does not satisfy the predetermined condition, that is, if the calculated error is not less than the allowable range (No), the process shifts to step S112 to execute the current control. A genetic operation is performed on the individual information used to construct the module 322. As the genetic operation here, for example, as shown in FIG.
Of the individual information currently used for the construction of the control module 322, a portion to which a synapse coupling coefficient in the neural network 322b is assigned (information that affects only the second evaluation value among the evaluation values) is targeted. Regenerate using random numbers. FIG. 8 is a diagram showing a target portion of the genetic operation and the genetic operation. Of course, after performing the genetic operation on the individual information, the control module 322 is reconfigured based on the new individual information. Then, the process proceeds to step S108, and until the predicted value of the response degree satisfies the predetermined condition, the process proceeds to step S1.
After repeating steps 08, S110, and S112, a genetic operation is performed on the individual information.

On the other hand, when it is determined in step S110 that the predicted value of the response degree satisfies the predetermined condition (Ye
In step s), the process shifts to step S114 to acquire the fuel efficiency and the response degree from the evaluation unit 330. Here, the control module 322 is constructed based on the individual information, and control of the engine 10 is started by the constructed control module 322. The fuel efficiency and the response obtained as a result are evaluation values for the individual. The higher the evaluation value, that is, the smaller the mileage as the first evaluation value and the higher the responsivity as the second evaluation value, the better the individual in the evolutionary simulation by GA. Can be positioned.

Next, the flow shifts to step S116, where CMAC3 is calculated based on the error between the measured value and the predicted value of the response.
Learning is performed on 42, and the process proceeds to step S118, where it is determined whether the processing from steps S106 to S116 is completed for all individuals in the individual group, and it is determined that the processing is completed for all individuals. Time (Yes)
Shifts to step S120.

In step S120, the fuel efficiency and the response level, which are the evaluation values for each individual, are displayed on the display unit 312 using the expression method shown in FIG. FIG. 9 (a)-
(C) is a figure which shows an example of the expression method of evaluation of an individual. In each figure, the width of the rectangle indicates the magnitude of the response, the color (shade) indicates the fuel efficiency, and the height of the rectangle is
The speed is shown divided into three stages. The responsivity increases in proportion to the width, the fuel efficiency improves in proportion to the color depth, and the upper speed indicates high speed, the middle speed indicates middle speed, and the lower speed indicates low speed. ing. The evaluation of each individual is performed for all individuals in the individual group, and is simultaneously displayed on one screen as shown in FIG. FIG. 10 is a diagram illustrating an example of a method of expressing individual evaluation.

Next, the process proceeds to step S 122, where the user's evaluation is input from the input unit 314. When the display of evaluations for all individuals in the population is completed, the control enters the evaluation mode once. In the evaluation mode, when the user selects an individual having characteristics that he / she wants to test-run by looking at the evaluation displayed on the display unit 312, the control module 322 is constructed based on the individual information of the individual selected by the user, and temporarily. It is fixed and the control by the control module 322 is performed. Thus, the user determines the characteristics of each individual displayed on the display unit 312 based on the actual riding comfort, and evaluates the evaluation value of each individual from the riding comfort. And
The process proceeds to step S124, and the user
When the evaluation of each individual based on the evaluation of the individual visually expressed and the evaluation of the individual comfort based on the actual running is completed, the control is switched to the selection mode, and the survival or selection of the individual in the population is performed. For the survival or selection of the individual, for example, the selection mode is switched to the input unit 314, and several individuals having the user's favorite characteristics are selected from the population while referring to a display screen as shown in FIG. This is performed by leaving the selected individual and erasing the other individuals.

Next, the flow shifts to step S126, where G
A crossover process for crossing individuals in A is executed. Specifically, in step S126, two parent individuals are selected from the population selected by the user using random numbers,
These are crossed to generate two offspring individuals. By performing this process five times, an individual group including nine child individuals is generated again (the tenth child individual is discarded). In the crossover process, for example, in addition to the above-mentioned one-point crossover process, 2
Point crossover processing or normal distribution crossover processing can be employed. The normal distribution crossover process is a process of generating child individuals according to a normal distribution that is rotationally symmetric with respect to an axis connecting the parent individuals, with respect to the individual information in the real-valued expression. The standard deviation of the normal distribution is such that the component in the main axis direction connecting the parents is proportional to the distance between the parents, and the components of the other axes are the third parent sampled from the straight line connecting the parents and the population. Make it proportional to the distance to the individual. This crossover method has the advantage that the characteristics of the parent individual are easily inherited by the child individuals.

Next, the flow shifts to step S128, where G
A mutation process for mutating the individual in A is executed, and the process proceeds to step S130, where it is determined whether or not characteristics that satisfy the user are obtained by the input from the input unit 314, and the user is satisfied. When it is determined that the characteristic cannot be obtained (No), the process proceeds to step S132, and it is determined whether or not the number of generation alternations is equal to or more than a predetermined number. Then, control goes to a step S134.

In step S134, an individual having the highest evaluation value is extracted from the individual group, an optimal control module is constructed based on the extracted individual information, and the control module 322 is fixed to the optimal control module. S
The process proceeds to 136, where the learning layer 200 learns the input / output relationship of the control module 322. The process proceeds to step S138, where the output of the control module 322 is set to “1”. Return to processing.

On the other hand, if it is determined in step S132 that the number of generation alternations is less than the predetermined number (Yes), the process proceeds to step S104. On the other hand, when it is determined in step S130 that the characteristics satisfying the user have been obtained (Yes),
The process moves to step S134. On the other hand, step S118
When it is determined that the processing from steps S106 to S116 has not been completed for all individuals in the population
(No) The process shifts to step S140 to read the individual information of the next individual in the individual group, and shifts to step S106.

Next, the operation of the above embodiment will be described with reference to the drawings. In order to optimize the control characteristics of the engine 10 for the user, the user first inputs an evolution start instruction to the input unit 314. In the evolution adaptation unit 320, when an evolution start instruction is input from the user, step S100,
Through S102, an individual group including a set of nine individuals is generated, and individual information is configured for each individual.

When the population is generated, the first generation evolution simulation is started. In the first generation evolution simulation, first, after step S106, the individual information of the first individual in the individual group is read, and the control module 322 is constructed and constructed based on the read individual information. The control of the engine 10 is started by the control module 322, and the control by the control module 322 is performed for a while. Meanwhile, step S1
Through steps 08, S110, and S112, a predicted value of the response degree is obtained from the prediction unit 340, and the individual information used for constructing the control module 322 is genetically manipulated until the obtained predicted value satisfies a predetermined condition. Is performed. In this genetic operation, the neural network 322 of the individual information currently used to construct the control module 322 is used.
Recombination by random numbers and regeneration by random numbers are performed only on the portion of b where the synapse coupling coefficient is assigned.

Next, when the predicted value of the responsivity satisfies the predetermined condition, the fuel consumption and the responsivity are obtained from the evaluation unit 330 through steps S114 and S116, and the CMAC 34 is calculated based on the error between the measured value and the predicted value of the responsivity.
The learning is performed on 2. In the same manner as described above, when the processing from step S106 to S116 is completed for all the individuals in the individual group, the fuel economy and the response level, which are the evaluation values, are displayed on the display unit 312 for each individual via step S120. . Here, the user selects several individuals having his / her favorite characteristics from the individual group while referring to the evaluation of each individual displayed on the display unit 312. When the user selects an individual, through step S124, the individual selected from the individual group is left, and the other individuals are deleted, whereby the individual survives or is eliminated.

Next, crossover processing and mutation processing are performed through steps S126 and S128. After the processing up to this point, the first generation evolution simulation ends. After that, the evolution simulation is repeatedly performed in the same manner until a characteristic satisfying the user is obtained or the number of generation alternations is equal to or more than a predetermined number.

Next, when the evolution simulation is completed, through step S134, an individual having the highest evaluation value is extracted from the population, and an optimal control module is constructed based on the extracted individual information of the individual. 322 is fixed to the optimal control module. Next, through step S136, the input / output relationship of the control module 322 is learned by the learning layer 200. In this learning, first, control is performed by using an evolutionary correction rate and an evolutionary transient correction rate for input information such as the actual engine speed obtained by the optimal control module. When the evolution adaptive layer 300 starts executing the control based on the evolution correction rate and the evolution transient correction rate, the learning neural network of the learning layer 200 changes the input / output relationship of the control module 322 to the learning layer 200.
Learning with the input / output relationship of the neural network functioning for the execution of During this time, the evolution adaptation layer 30
The output of 0 is performed by the individual whose evaluation function before that is maximized, and the control law does not change with time. In the learning described above, inputs and outputs between the evolution adaptive layer 300 and the neural network for execution of the learning layer 200 are averaged with a certain step width, and the average is used as input / output data for updating the teacher data set. For example, the average engine speed per second is 5000 [rpm], the average throttle opening is 20,
The average intake air temperature is 28 [° C] and the average atmospheric pressure is 1013 [h
Pa], these and the evolutionary adaptation layer 3 at that time
00 and the output of the neural network for execution in the learning layer 200 are used as input / output data (see FIG. 11). This input / output data is added to the previous teacher data to obtain a new teacher data set. At this time, old teacher data whose Euclidean distance with new data in the teacher data set is within a certain value is deleted. This is shown in FIG. The initial value of the teacher data set is set to “1” for all input data. The learning layer 200 learns synaptic coupling coefficients in the learning neural network based on the updated teacher data set. The learning of the coupling coefficient is based on the output of the learning neural network during learning and the reflection layer 100.
Is performed until the error between the virtual control output obtained from the basic injection amount and the transient correction rate and the actual control output becomes equal to or smaller than the threshold, and when this learning is completed, the neural network for learning is This is for execution, and the original neural network for control is for learning. Thereafter, the learning layer 200 determines the learning correction rate and the learning transient correction rate using the newly obtained execution neural network and actually outputs them. At the same time, after step S138, the output of the control module 322 becomes “1”. And the learning layer 20
0 and the reflection layer 100 are controlled. The initial value of the neural network for execution of the learning layer 200 is set so that the output is always “1”. By doing so, in the initial state, control can be performed only by the reflective layer 100 and the evolution adaptive layer 300.

As described above, in the present embodiment, the evolution adaptation unit 320 predicts the responsivity for each individual based on the individual information in the evolution simulation by the GA, and sets the predicted responsivity to a predetermined condition. Genetic operations were performed on the individual information until the condition was satisfied. This further increases the possibility that the evolution of the population will be promoted in one generation by repeating the prediction and the genetic operation based on the prediction, so that it is further expected that an effective solution can be obtained with a small number of generation alternations. Therefore, optimization of control characteristics can be realized in a relatively short time.

Further, in this embodiment, a genetic operation is performed on the synapse coupling coefficient in the neural network 322b in the individual information based on the predicted value of the response level. Thus, it can be expected that the evolution will be promoted specializing in the improvement of the response level.

Further, in the present embodiment, the first evaluation value is the fuel efficiency of the engine 10, and the second evaluation value is the degree of response. As a result, the population evolves in a direction to improve the fuel efficiency and the response level.
Among the control characteristics of 0, the fuel consumption characteristics and the response characteristics are optimized. In particular, with regard to response characteristics, it is expected that repetition of prediction and genetic operations based on them will further increase the possibility of promoting population evolution in one generation, so it is expected that effective response characteristics will be obtained with a small number of generation alternations it can.

Further, in this embodiment, the engine 10
The correction rate of the fuel injection amount or the correction rate of the transient correction rate is generated by the neural networks 322a and 322b, and the synapse coupling coefficient in the neural networks 322a and 322b is assigned to the individual information. I have. As a result, it is expected that a high evaluation value can be obtained.
The coupling coefficient of the synapse at 2a and 322b can be determined.

In the above embodiment, the engine 10
Corresponds to the control target according to any one of claims 6 to 10, 12 and 14, and the prediction unit 340 performs the control according to claims 3, 5 to 7,
The synapse coupling coefficient in the neural networks 322a and 322b corresponds to the control coefficient according to claim 6, 10 or 12. Step S112 corresponds to the individual information operation means described in claim 6, and the evaluation section 330 corresponds to the evaluation value calculation means described in claim 6.
Step S124 corresponds to the individual selecting means described in claim 6.

In the above embodiment, the vehicle engine 10 is applied as a control target of the control device 30. However, the control target of the control device 30 is not limited to the present embodiment, but is arbitrary. For example, control of suspension characteristics of a vehicle body or damper characteristics of a seat, assist characteristics of an auxiliary power in a bicycle or a wheelchair using an electric motor or an engine as an auxiliary power, or operation characteristics of a personal robot (a crisp operation or a leisurely operation) Operation) may be applied.

Further, in the present embodiment, the fuel injection amount is treated as the control output. However, when the engine 10 is applied as a control target, the control output may include, for example, injection time, ignition timing, The intake valve timing, the electronic throttle opening, the valve lift, the exhaust valve timing, the intake / exhaust control valve timing, and the like can be considered. Here, the intake control valve is a valve provided on the intake pipe for controlling tumble and swirl, and the exhaust control valve is a valve provided on the exhaust pipe for controlling exhaust pulsation. is there.

In the present embodiment, the learning layer 2
Although 00 is configured as a hierarchical neural network, the configuration of the control system of the learning layer 200 is not limited to this embodiment, and for example, CMAC may be used. CMA
Advantages of using C include superior additional learning ability and high-speed learning as compared with the hierarchical neural network.

In the above embodiment, the individual information of each individual includes the neural networks 322a and 322a.
Although the configuration is such that the synapse coupling coefficient in 22b is assigned, the present invention is not limited to this, and a variation as shown in FIG. 13 can be considered. FIG. 13 is a diagram illustrating another data structure of the individual information. In FIG. 13A, a coefficient for constructing a fuel control module is assigned to individual information. In FIG. 13B, respective coefficients for constructing the electronic throttle module and the fuel control module are assigned to individual information. In FIG. 13C, the electronic throttle module, the fuel control module and the C
Each coefficient for constructing a VT (Continuously Variable Transmission) module is assigned to individual information.

In the above embodiment, the prediction unit 340 uses the CMAC 342 as a learning mechanism.
Instead, a neural network may be used as the learning mechanism. Further, in the above embodiment, the individual is selected based on the evaluation by the user. However, the present invention is not limited to this, and the evolution adaptation unit 320 automatically performs the selection based on a predetermined criterion. May be configured.

Further, in the above embodiment, the correction rate of the fuel injection amount of the engine 10 or the correction rate of the transient correction rate is generated by the neural networks 322a and 322b. The configuration is such that the synapse coupling coefficient in 322a and 322b is assigned. However, the present invention is not limited to this. The individual information is directly assigned with the correction rate of the fuel injection amount of the engine 10 or the correction rate of the transient correction rate. Is also good.

As a result, it is possible to determine the correction rate of the fuel injection amount of the engine 10 or the correction rate of the transient correction rate at which a high evaluation value can be expected. In the above embodiment, the correction rate of the fuel injection amount of the engine 10 or the correction rate of the transient correction rate is generated by the neural networks 322a and 322b.
The present invention is not limited to this, and the neural network 322a, 322b may be configured to generate the fuel injection amount, the transient correction amount, the correction amount of the fuel injection amount, or the correction amount of the transient correction amount of the engine 10. This is the same for the configuration in which the calculation is performed directly without being generated by the neural networks 322a and 322b.

In the above embodiment, the GA is used to optimize the fuel efficiency and response characteristics of the engine 10 for the user. However, the present invention is not limited to this, and evolutionary algorithms such as GP and ES may be used. It can also be used. In the above embodiment, the interface unit 3
10 based on the evaluation by the user.
However, the control characteristics of the engine 10 are evaluated based on a predetermined evaluation criterion, and the control characteristics of the engine 10 are autonomously optimized based on the evaluation. May be configured. That is, in implementing the present invention, the individual evaluation may be an interactive evaluation or an autonomous evaluation.

In the above embodiment, when executing the processing shown in the flowchart of FIG.
The case where a program stored in advance in M is executed has been described. However, the present invention is not limited to this. For example, the program may be read from a recording medium on which a program indicating these procedures is recorded and executed by being read into the RAM. Good.

Here, the recording medium is a semiconductor recording medium such as a RAM or a ROM, a magnetic recording type recording medium such as an FD or HD, an optical reading type recording medium such as a CD, CDV, LD, or DVD; It is a magnetic recording type / optical reading type recording medium, and includes any recording medium that can be read by a computer irrespective of an electronic, magnetic, optical or other reading method.

[0101]

As described above, according to the optimal solution searching apparatus by the evolutionary method according to the first to fifth aspects of the present invention, the evolution of the population in one generation by the prediction and the genetic operation based on the prediction. Therefore, there is an effect that it is possible to expect to obtain an effective solution with a smaller number of generation alternations than in the past.

Further, according to the optimal solution search device by the evolutionary method according to the second aspect of the present invention, there is a possibility that the evolution of the population in one generation is promoted by repeating the prediction and the genetic operation based thereon. Since the cost is higher, it is possible to obtain an effect that a more effective solution can be expected with a smaller number of generation alternations. On the other hand, according to the control device of the controlled object by the evolutionary method according to claims 6 to 14 of the present invention, the evolution of the population in one generation may be promoted by the prediction and the genetic operation based on the prediction. It can be expected that an effective solution can be obtained with a smaller number of generation alterations than in the past. Therefore, an effect is obtained that the control characteristics can be optimized in a relatively short time.

Further, according to the control device for the controlled object by the evolutionary method according to the seventh aspect of the present invention, there is a possibility that the evolution of the population in one generation is promoted by repeating the prediction and the genetic operation based thereon. Is even higher,
It is further expected that a valid solution can be obtained with a small number of generation alternations. Therefore, an effect is obtained that the control characteristics can be optimized in a shorter time.

Further, according to the control device of the controlled object by the evolutionary method according to the eighth aspect of the present invention, an effect is obtained that the evolution can be expected to be accelerated specializing in the improvement of the second evaluation value. . Furthermore, according to the control device for the control object by the evolutionary method according to the ninth aspect of the present invention, since the population evolves in a direction to improve the fuel efficiency and the response, the fuel efficiency characteristic among the engine control characteristics is improved. And the response characteristics are optimized. In particular, with regard to response characteristics, it is expected that repetition of prediction and genetic operations based on them will further increase the possibility of promoting population evolution in one generation, so it is expected that effective response characteristics will be obtained with a small number of generation alternations The effect that can be obtained is also obtained.

Further, according to the control device for the control object by the evolutionary method according to the eleventh aspect of the present invention, it is expected that a high evaluation value can be obtained.
The effect that the correction value of the fuel injection amount or the correction value of the transient correction amount can be determined is also obtained. Further, according to the control device of the control object by the evolutionary method according to claim 12 of the present invention, it is expected that a high evaluation value is obtained.
There is also obtained an effect that the coupling coefficient of the synapse in the neural network can be determined.

Further, according to the control device for the control object by the evolutionary method according to the thirteenth aspect of the present invention, the population evolves in a direction to improve the power consumption and the rate of change in rotation. The power consumption characteristics and the rotation change characteristics among the control characteristics are optimized. In particular, as for the rotation change characteristics, it is more likely that the evolution of the population will be promoted in one generation by repeating the prediction and the genetic operation based on it, so it is necessary to obtain an effective rotation change characteristic with a small number of generation alternations Can be expected.

On the other hand, according to the optimal solution search program by the evolutionary method according to claim 15 of the present invention, claim 1
An effect equivalent to that of the optimal solution search device by the described evolutionary method is obtained.

[Brief description of the drawings]

FIG. 1 is a diagram showing a basic concept of the present invention.

FIG. 2 is a block diagram showing a basic configuration of the present invention.

FIG. 3 is a block diagram showing a configuration of an engine control system to which the present invention is applied.

FIG. 4 is a block diagram showing a configuration of an evolution adaptive layer 300.

FIG. 5 is a diagram showing regions used in a reference individual.

FIG. 6 is a diagram showing a configuration of neural networks 322a and 322b and a data structure of individual information.

FIG. 7 is a flowchart showing processing executed by an evolution adaptation unit 320.

FIG. 8 is a diagram showing a target portion of the genetic operation and the genetic operation.

FIG. 9 is a diagram showing an example of a method of expressing individual evaluation.

FIG. 10 is a diagram showing an example of a method of expressing individual evaluation.

FIG. 11 is a diagram conceptually showing a state in which a teacher data set acquires new teacher data.

FIG. 12 is a diagram conceptually showing updating of a teacher data set.

FIG. 13 is a diagram showing another data structure of individual information.

[Explanation of symbols]

 Reference Signs List 10 engine 20 various sensors 30 control device 100 reflection layer 110 basic control unit 200 learning layer 210 learning unit 300 evolution adaptation layer 310 interface unit 312 display unit 314 input unit 320 evolution adaptation unit 322 control module 322a, 322b neural network 330 evaluation unit 332 Fuel consumption calculation unit 334 Response degree calculation unit 340 Prediction unit 342 CMAC

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI theme coat ゛ (Reference) G05B 13/02 G05B 13/02 LZ F-term (Reference) 3G084 BA13 DA02 DA07 EB17 EC04 FA10 FA13 FA30 FA33 3G301 JA02 LB01 LB11 MA11 NA09 ND25 ND30 ND42 ND43 PA11Z PA12Z PB03A PE01Z 5H004 GA18 GB12 KC50 KD33 KD43 KD67

Claims (15)

[Claims] 1. A group of a plurality of individuals is virtually generated, and individual information is configured for each of the individuals based on the genetic information of the individual, and genetic manipulation is imitated in the same generation. A genetic operation for performing information operation on the individual information, and an individual selection operation for performing survival or selection of the individual based on an evaluation value of the individual obtained by an evaluation function are respectively performed at least once, and the generation proceeds. An apparatus for searching for an optimal solution of the evaluation function, wherein the apparatus estimates the evaluation value and performs the genetic operation based on the predicted value. Optimal solution search device. 2. The optimal solution search device according to claim 1, wherein the genetic operation and the prediction are repeatedly performed until the predicted value satisfies a predetermined condition. 3. The predicting unit according to claim 1, wherein the evaluation value includes a first evaluation value and a second evaluation value, and a prediction unit that predicts the second evaluation value based on the individual information. An optimal solution search device based on an evolutionary method, wherein the genetic operation and the prediction are repeatedly performed until a predicted value predicted by the prediction unit satisfies a predetermined condition. 4. The method according to claim 3, wherein the genetic operation is performed on information that affects the second evaluation value among the individual information, based on the predicted value. Solution search device based on evolving evolutionary method. 5. The method according to claim 3, wherein the prediction unit predicts the evaluation value by learning, and performs learning based on a difference between the prediction value and the evaluation value. An optimal solution search device using an evolutionary method characterized by the following. 6. An individual group consisting of a set of a plurality of individuals is virtually generated, and individual information is configured for each individual based on genetic information of the individual, and the individual information includes a control target Assigning a control coefficient that influences a control characteristic of a control system that controls the characteristic, furthermore, an individual information operating means for performing an information operation imitating a genetic operation on the individual information, and an evaluation for calculating an evaluation value of the individual Value calculation means, and individual selection means for performing survival or selection of the individual based on the evaluation value calculated by the evaluation value calculation means, and in the same generation, the genetic operation by the individual information operation means and the individual selection. Means for performing the individual selection operation at least once by each means to advance the generation, and using the individual information of the individual evolved thereby as the control coefficient, A device for optimizing the control characteristics of the device, comprising a prediction unit for predicting the evaluation value based on the individual information, and performing the genetic operation based on the prediction value predicted by the prediction unit. A control device for an object to be controlled by an evolutionary method characterized in that: 7. The evaluation value according to claim 6, wherein the evaluation value includes a first evaluation value and a second evaluation value, and the prediction unit predicts the second evaluation value based on the individual information. A control apparatus for controlling a control target by an evolutionary method, wherein the genetic operation and the prediction are repeatedly performed until a predicted value predicted by the prediction unit satisfies a predetermined condition. 8. The method according to claim 7, wherein the genetic operation is performed on information that affects the second evaluation value in the individual information based on the predicted value. A control device for a controlled object by an evolutionary method. 9. The engine according to claim 7, wherein the control target is an engine, the first evaluation value is a fuel efficiency of the engine, and the second evaluation value is a rotation speed of the engine. A control device according to an evolutionary method, characterized in that the response is determined by a change rate and a throttle opening change rate. 10. The fuel injection amount of the engine and the fuel in the transient state of the engine obtained when the engine is operated by using the individual information as the control coefficient. A control device according to an evolutionary method, wherein the second evaluation value is predicted based on a transient correction amount for correcting an injection amount. 11. The engine according to claim 6, wherein the control target is an engine, and the individual information includes the control element as the control coefficient, a fuel injection amount of the engine, and the fuel in a transient state of the engine. A control device according to an evolutionary method, wherein a transient correction amount for correcting an injection amount, a correction value for the fuel injection amount, or a correction value for the transient correction amount is assigned. 12. The engine according to claim 6, wherein the control target is an engine, a fuel injection amount of the engine, a transient correction amount for correcting the fuel injection amount in a transient state of the engine, and the fuel. The correction value of the injection amount or the correction value of the transient correction amount is generated by a neural network, and the individual information is assigned a coupling coefficient of a synapse in the neural network as the control coefficient. A control device for an object to be controlled by an evolutionary method characterized in that: 13. The electric motor according to claim 7, wherein the control target is an electric motor, the first evaluation value is power consumption of the electric motor,
The control device according to an evolutionary method, wherein the second evaluation value is a rotation change rate of the electric motor. 14. The method according to claim 6, wherein the prediction unit predicts the evaluation value by learning and performs learning based on a difference between the predicted value and the evaluation value. A control device for an object to be controlled by an evolutionary method characterized in that: 15. An individual group consisting of a set of a plurality of individuals is virtually generated, and individual information is configured for each of the individuals based on the genetic information of the individual, and genetic manipulation is imitated in the same generation. A genetic operation for performing information operation on the individual information, and an individual selection operation for performing survival or selection of the individual based on an evaluation value of the individual obtained by an evaluation function are respectively performed at least once, and the generation proceeds. A computer-executable program for searching for an optimal solution of the evaluation function, wherein the genetic value is predicted based on the predicted value and the genetic operation is performed based on the predicted value. Optimal solution search program by evolving evolutionary method.