JP2000250604A - Cooperation method of optimization for characteristic optimization method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To secure the optimum cooperation in order to attain a desired purpose despite the trade-off relation confirmed between the optimization results by optimizing a control module used for the normal control and then optimize another control module used for the normal control to improve or maintain the acquired characteristic. SOLUTION: When plural controlled systems exist and every control module is optimized, an optimization processing part is prepared for every control module that is evolved. Under such conditions, some characteristic changes are affected with each other according to the relation of the controlled systems and a characteristic optimization is sometimes traded off. In such a case, the cooperative optimization processing is secured among plural control modules. For example, after one control module is optimized, another control module is optimized so as to improve or maintain the acquired characteristic. Thus, it's possible to optimize the characteristic of another controlled system in the optimum characteristic range of one controlled system.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for optimizing the characteristics of a control module for controlling an object to be controlled.

[0002]

2. Description of the Related Art Conventionally, an optimum value of a characteristic of a control module for controlling a control target (that is, a value of a parameter which determines an input / output relationship of the control module) is assumed for a user of a product to be controlled. However, it has been determined by experiment in the design stage or the setting stage before shipment so as to match the assumed characteristics (preference, skill, personality, use state) of the user. However, with the recent diversification and sophistication of control contents, the conventional method of deciding the optimum value of the characteristics of the control module by experiment increases the difficulty of optimizing the control module, and therefore takes a lot of time. Is required. In view of the conventional problems described above, the applicant optimizes, using a genetic algorithm, the characteristics of a control module for normal control that determines an output related to an operation amount of a control target based on predetermined input information. Japanese Patent Application No. Hei 9 for optimizing characteristics
-264604. According to this method, optimization can be performed in a short time by using a genetic algorithm for optimizing characteristics.

[0003]

However, in the characteristic optimization method proposed by the applicant, a plurality of characteristics such as fuel consumption performance and drivability performance are optimized for one control object such as a fuel injection device. However, the problem remains that if the fuel efficiency is improved, the drivability is reduced, and if the drivability is improved, the fuel efficiency is reduced. I was Further, when there are a plurality of control modules for normal control that determine the output related to the manipulated variable of the control target, a similar problem remains when optimizing the characteristics of each control module for normal control. The present invention solves the problems of the conventional optimization method described above, and optimizes a plurality of characteristics.
It is an object of the present invention to provide a method of coordinating the results of optimization of both characteristics so as to be optimal in order to achieve a target in a trade-off relationship.

[0004]

In order to achieve the above object, an optimization coordination method in the characteristic optimization method according to the first aspect of the present invention provides a method of operating a control object based on predetermined input information. In a characteristic optimization method for optimizing the characteristics of a plurality of control modules for normal control, each of which determines an output related to the control module, after optimizing one control module for normal control, the acquired characteristics are improved or maintained. Optimize other normal control modules. Further, the optimization coordination method in the characteristic optimization method according to claim 2 of the present invention is characterized in that the characteristic of a plurality of normal control control modules for determining an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimization method for optimizing
The plurality of control modules are optimized at regular intervals so as to improve or maintain the characteristics acquired from each other. Also,
The optimization coordination method in the characteristic optimization method according to claim 3 of the present invention is characterized in that each of the characteristics of a plurality of normal control control modules for determining an output related to an operation amount of a control target based on predetermined input information is determined. In the characteristic optimization method to be optimized, while optimizing one normal control control module, optimizing another normal control control module so as to improve or maintain the characteristics acquired by the normal control control module. Are performed in parallel. The optimization coordination method in the characteristic optimization method according to claim 4 of the present invention is characterized in that a characteristic of a plurality of normal control control modules for determining an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimization method for optimizing each of the control parameters, a plurality of control modules for normal control are optimized in parallel so as to improve or maintain the characteristics obtained by each other. Furthermore, claim 5 of the present invention
The coordination method for optimization in the characteristic optimization method according to the present invention is based on the coordination method according to any one of claims 1 to 4, based on an evaluation criterion preset for optimization of at least one control module for normal control. An interactive evaluation method that uses an autonomous evaluation method that performs evaluation during optimization processing and performs evaluation during optimization processing based on evaluation based on the user's will to optimize other control modules for normal control. use. Also,
When the evaluation criterion is set in advance, for example, the evaluation criterion can be set based on the reference characteristic of the control target accompanied by the deterioration with time or the regulation on the control target. Further, if necessary, for example, an evaluation standard based on the regulation is set in advance for one control object, and a construction is made so that the evaluation based on the user's will can be performed within the range of the evaluation standard. The evaluation may be used in combination such as optimizing the characteristics according to the user's preference within the range of the above. In addition, the optimization coordination method in the characteristic optimization method according to claim 8 of the present invention is characterized in that at least one normal control control module that determines an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimization method for optimizing a plurality of characteristics, after optimizing one characteristic, other characteristics are optimized so as to improve or maintain the obtained characteristic. In addition, the optimization coordination method in the characteristic optimization method according to claim 9 of the present invention provides a method for controlling a plurality of characteristics of a control module for normal control that determines an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimization method for optimizing each of the
Optimization is performed at regular intervals so as to improve or maintain the characteristics acquired by each other. In addition, the optimization coordination method in the characteristic optimization method according to claim 10 of the present invention provides a method for controlling a plurality of characteristics of a control module for normal control that determines an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimization method for optimizing each characteristic, while optimizing one characteristic, optimization of another characteristic is performed in parallel so as to improve or maintain the characteristic. In addition, the optimization coordination method in the characteristic optimization method according to claim 11 of the present invention includes a plurality of characteristics of a control module for normal control that determines an output related to an operation amount of a control target based on predetermined input information. Are optimized in parallel so as to improve or maintain mutually acquired characteristics. Further, claim 12 of the present invention
The optimization coordination method in the characteristic optimization method according to the present invention is the coordination method according to any one of claims 6 to 10, wherein:
An autonomous evaluation method that performs an evaluation during the optimization process based on an evaluation criterion set in advance for optimization of at least one normal control control module, and a user's will to optimize other normal control control modules We use an interactive evaluation method that evaluates during the optimization process based on the evaluation based on. When the evaluation criterion is set in advance, for example, the evaluation criterion may be set based on the reference characteristics of the control target with time-dependent deterioration or the regulation on the control target. Further, if necessary, for example, for one control target,
Evaluation criteria based on the regulation are set in advance, and built so that the evaluation based on the user's will can be performed within the range of the evaluation criteria, and the characteristics can be optimized according to the user's preference within the regulation. And evaluation may be used in combination.

[0005]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a characteristic optimization method according to the present invention; The form will be described. 1 to 13
Shows an embodiment in which the characteristic optimization method employing the coordination method according to the present invention is applied to the control of a vehicle engine and a continuously variable transmission. FIG. 1 is a schematic diagram showing a relationship between an engine 1 and a control device 10 that executes a comprehensive characteristic optimizing method.
The control device 10 is configured to improve fuel efficiency while obtaining desired drivability and acceleration.
In this specification, “drivability” means the response of the engine output to the throttle operation. As shown in the drawing, the control device 10 inputs information such as an engine speed, an intake negative pressure, an accelerator operation amount, an atmospheric pressure, an intake air temperature, a cooling water temperature, and the like, based on the input information, a fuel injection device, an electronic throttle, and the like. By operating the valve and the continuously variable transmission, the amount of fuel injection, the amount of intake air, and the gear ratio are controlled, and optimal control that achieves both drivability, acceleration, and fuel efficiency is performed. FIG. 2 is a schematic block diagram of the control device 10. As shown in the drawing, the control device 10 has an electronic throttle control unit, a continuously variable transmission control unit, and a fuel injection device control unit. An electronic throttle control module that determines an opening of the electronic throttle valve based on predetermined input information;
An optimization processing unit that optimizes control parameters of the electronic throttle control module. The continuously variable transmission control unit determines a basic speed ratio of the continuously variable transmission based on predetermined input information (outer world information in FIG. 2), and determines a correction rate for the basic speed ratio. And a optimizing unit for optimizing the correction module. The fuel injection device control unit includes a fuel injection device control module that determines a basic fuel injection amount based on predetermined input information (outside information in FIG. 2), and a fuel injection amount correction that determines a correction rate for the basic fuel injection amount. Module, an optimization processing unit that optimizes the correction module, and an evaluation unit that evaluates the optimization calculation unit.

As shown in FIG. 3, the electronic throttle control module determines the opening of the electronic throttle valve based on the accelerator operation amount of the user. It should be noted that the “accelerator operation amount” in this specification includes both the information of the actual “accelerator angle” and the information of the “accelerator change amount”. Here, the characteristics of the electronic throttle valve will be briefly described. The electronic throttle valve has two characteristics of a static characteristic and a dynamic characteristic. The former is a characteristic resulting from the relationship between the accelerator angle and the electronic throttle valve, and affects the steady running characteristics of the vehicle. FIG. 4 is a graph showing examples of static characteristics of some throttles. By changing the static characteristics in this way, when the accelerator angle is small, the electronic throttle valve opens widely, and as the accelerator angle increases, the throttle valve gradually converges to full opening. The throttle valve opens gradually while it is small, and when the accelerator angle becomes large, it suddenly opens to the full throttle.The high opening rapid acceleration type and the proportional type where the accelerator angle and throttle opening are proportional are various at the same accelerator angle depending on the setting. It is possible to obtain a proper throttle opening. The static characteristics need only be such that the throttle opening increases or does not change as the accelerator angle increases, and various functions can be obtained. In this embodiment, the throttle opening is 0 to 2
Throttle valve opening ratio SP1 at 0% and throttle valve opening ratio SP2 at throttle opening of 20 to 100%
By optimizing the above, the static characteristics are optimized. Also, the latter characteristic of the electronic throttle valve, that is, the dynamic characteristic is
This characteristic is generated from the speed of change of the throttle valve with respect to the speed of change of the accelerator, and affects the transient characteristics of the vehicle. Specifically, this characteristic can be configured so that the change speed of the throttle with respect to the change speed of the accelerator can be changed by combining the first-order lag and the incomplete differentiation. By combining the first-order lag and the imperfect differentiation in this manner, as shown in FIG. 5, a type in which the throttle opens relatively slowly with respect to the accelerator operation, and a low spike occurs with respect to the accelerator operation, Various dynamic characteristics can be obtained, such as a type that responds quickly and has a high response that the throttle opens, or a type that is intermediate between the two.
In this embodiment, the dynamic characteristics are optimized by optimizing the first-order lag time constant DR and the acceleration correction coefficient AG. The optimization processing unit in the electronic throttle control unit employs, for example, an evolutionary calculation method as the optimization operation, and controls the control parameters (throttle valve opening ratios SP1 and SP2,
The temporary delay time constant DR and the acceleration correction coefficient AG) are coded as one individual as shown in FIG. 6, and these control parameters are optimized using an evolutionary calculation method. The evaluation of the value of each control parameter during the optimization process is configured to be performed based on the drivability that the user actually experiences. As a result, each control parameter in the electronic throttle control module is evaluated by the user. , And optimal electronic throttle characteristics (drivability characteristics) suitable for the user's evaluation can be obtained. Such a method in which the user performs the evaluation in the optimization processing is referred to as an interactive evaluation in this specification. In the present embodiment, the static characteristics and the dynamic characteristics are combined into one individual to optimize the entire combination. However, there are some other methods as described below. 1. The driver sets the static characteristics in advance, and optimizes only the dynamic characteristics. 2. Static and dynamic characteristics are optimized independently and separately. 3. The static characteristics are first evolved and fixed, and the dynamic characteristics are optimized.

[0007] The continuously variable transmission control module outputs a basic gear ratio for predetermined input information (for example, vehicle speed and throttle valve opening) based on a basic gear ratio map. The speed ratio correction module is configured by a neural network that outputs a correction ratio for the basic speed ratio based on predetermined input information (for example, vehicle speed and throttle valve opening). The optimization processing unit in the continuously variable transmission control unit includes:
As the optimization operation, for example, an evolutionary calculation method is adopted, an individual is generated by coding a coupling coefficient (control parameter) of a neural network constituting the transmission ratio correction module, and the evolutionary calculation method is used. Then, these coupling coefficients (control parameters) are optimized. The evaluation of the value of each control parameter during the optimization process is configured to be performed based on the acceleration feeling actually experienced by the user, and as a result, the control parameters of the gear ratio correction module are changed by the user. The optimum continuously variable transmission characteristics (acceleration) that are optimized according to the evaluation and match the user's evaluation are obtained.

[0008] The fuel injection device control module includes, for example,
As shown in FIG. 7, the basic fuel injection amount of the fuel injection device is determined based on the forward model of the engine modeled using the feedforward control logic with the learning function, and the output of the forward model and the target air-fuel ratio. And a fuel injection amount determining unit that performs the determination. The target air-fuel ratio is calculated by a target air-fuel ratio calculator based on the engine speed and the throttle opening. FIG. 8A shows a fuel injection amount correction module.
As shown in the figure, it is composed of a neural network that inputs the throttle opening and the engine speed and outputs a correction rate, and the obtained correction rate is multiplied by the basic fuel injection amount output from the fuel injection device control module. Thus, the final fuel injection amount is obtained. The optimization processing unit in the fuel injection device control unit employs, for example, an evolutionary calculation method as an optimization operation, and as shown in FIG. 8B, a neural network constituting a fuel injection amount correction module. Coding coefficients are coded to generate individuals, and these coupling coefficients (control parameters) are calculated using an evolutionary calculation method.
Perform optimization. The evaluation of the value of each control parameter during the optimization process is configured to be performed by the evaluation unit in which the target fuel efficiency is set. As a result, the coupling coefficient (control parameter) of the fuel injection amount correction module is The fuel efficiency is automatically optimized for the target fuel efficiency, and the optimum fuel efficiency characteristics can be obtained. In this manner, a method for performing the evaluation in the optimization operation by an evaluation unit designed in advance and performing the optimization automatically is referred to as an autonomous evaluation in this specification.

Next, the optimization process in the electronic throttle control unit, the continuously variable transmission control unit, and the fuel injection device control unit will be described. FIG. 9 is a flowchart showing the flow of the optimization processing of the entire control device 10. As described above, when performing the optimization processing, the control device 10 uses the interactive evaluation for the electronic throttle control unit and the continuously variable transmission control unit, and uses the autonomous type for the fuel injection device control unit. I have. Since the flow of the optimization process differs depending on the evaluation method, the following description will be made separately on the optimization process using the interactive evaluation method and the optimization process using the autonomous evaluation method.

A. Optimization Process in Electronic Throttle Control Unit and Continuously Variable Transmission Control Unit As shown in FIG. 9, first, control parameters to be optimized by the control module to be optimized in each control unit (in the case of the electronic throttle control module, , Static characteristics SP
1, SP2, dynamic characteristics DR and AG. In the case of a gear ratio correction module, the initial value of the coupling coefficient of the neural network constituting it is randomly determined within a predetermined range, and a plurality of initial values are determined. A first generation consisting of individuals is generated (step 1-1). Then, a test drive is performed using the parameters of any individual of the first generation (step 1-2), and the user inputs an evaluation value for the individual (step 1-).
3). Based on the evaluation value, it is determined whether or not the desired drivability or acceleration is obtained (step 1-4). If it is determined that the drivability or acceleration is obtained, the evolution process is terminated. Determines whether the test drive and evaluation have been completed for all individuals of one generation (step 1-5).
If the test drive and evaluation have not been completed for all individuals, the parameters of the control module are changed to those of another individual (step 1-6), and the test drive is performed again (step 1-2). When the test drive and evaluation for all the individuals have been completed, the process enters the evolutionary calculation module (step 1-7), generates a next-generation population, and again performs test drive and test using the parameters of those individuals. Perform an evaluation. These processes are repeated until the desired drivability or acceleration is obtained. As a result, the parameters of the electronic throttle control module and the gear ratio correction rate module are optimized. Here, a description will be given of the evaluation of drivability and acceleration using an interactive type. An input device of an evaluation value that can be operated during driving is realized by a button or the like so that a user can intervene in evolution. The user presses this button after the test drive to input the evaluation value of the individual who took the test drive. The evaluation value is determined based on the length of time the button is pressed.
Specifically, for example, there is a method of multiplying the reciprocal of the pressed time by a constant coefficient or a method of calculating using a fuzzy rule. By doing so, even if there is ambiguity in human evaluation, an evaluation value can be obtained with a certain degree of accuracy, and a user can interactively use the evolution method. Further, when the button is pressed for more than a predetermined time, the individual under evaluation at that time can be switched to the next individual. By doing so, the user can immediately change an individual who does not like the characteristic, and can evolve at high speed. Switching of individuals is performed only when the vehicle is stopped. This is effective for eliminating the effect of a sudden change in the throttle characteristic during traveling.

B. Optimization Process in Fuel Injection Control Unit As shown in FIG. 9, first, control parameters to be optimized by a control module to be optimized in each control unit (in the case of a fuel injection amount correction module, it is constructed. The initial value of the coupling coefficient of the neural network is randomly determined within a predetermined range to generate a first generation including a plurality of initial individuals (step 2-1). Perform fuel economy calculation (Step 2-
2). Here, the fuel consumption calculation will be briefly described. In the fuel injection control evolution module, a plurality of individuals are operated in a pseudo-parallel manner by time division, and the evaluation values in the total of the periods are compared. Specifically, for example, as shown in FIG. 10, control is performed for one minute at a time for ten individuals, and this is defined as one cycle, and the cycle is repeated for 20 cycles, and the total mileage during the evaluation period is divided by the fuel consumption. To calculate the fuel efficiency, that is, the evaluation value. By doing so, the effect of the difference in gear position and the effect of the climbing angle can be adjusted as a total for each individual, so that the characteristics of each individual can be evaluated fairly. Based on the fuel efficiency (that is, the evaluation value) of each individual obtained in the above-described fuel efficiency calculation process (step 2-2), it is evaluated whether or not it is an optimum fuel efficiency characteristic (step 2-3). As a result, it is determined whether or not the optimum fuel efficiency has been obtained (step 2-4). If the optimum fuel efficiency has been obtained, the optimization process is terminated. If the optimum fuel efficiency has not been obtained, the process enters the evolutionary calculation module (step 2-5) to generate the next generation population.

Here, some examples of the evolutionary computing module will be described. a. Genetic Algorithm (GA) FIG. 11 is a schematic flowchart of an evolutionary calculation module when a genetic algorithm is used as an evolutionary calculation method. In this module, after the evaluation of all the individuals of one generation is completed, if a desired characteristic is not obtained, a population of the next generation is generated. As for the scaling (step 1), the fitness is linearly converted so that the ratio between the maximum fitness and the average fitness in the individual group is constant. As for the selection (step 2), a roulette selection method in which the selection is established in proportion to the evaluation value (fitness) of the user can be adopted. Alternatively, a tournament selection method of selecting a randomly selected n individuals having the best evaluation value may be used. Crossover (step 3) has one point crossover, 2
There are methods such as point crossover and normal distribution crossover. It is possible that the selected crossover parent is the same fixed, but if left unchecked, diversity as a population is lost. Therefore, if the parents selected for crossover are the same individual,
Swap with other selected individuals to avoid crossover of the same individuals as much as possible. For the mutation (step 4), the value is changed randomly at a certain probability for each locus of the individual. In addition, a method of adding a perturbation according to a normal distribution is also conceivable. If different individuals are selected as crossover parents, but they are genetically identical, then both crossing parents will be mutated with a higher probability than normal. In addition to the above, a method of generation alternation called “regeneration” that replaces all individuals of one generation at a time can also be used. In addition, if strict generational changes are applied, individuals with high evaluations may be destroyed. Therefore, use an elite conservation strategy that unconditionally leaves the elite (individuals with the highest evaluation) in the next generation. Can be.

B. Evolution Strategy (ES) FIG. 12 is a schematic flowchart of an evolutionary computation module when an evolutionary strategy is used as an evolutionary computation method. In this module, after all the individuals of one generation have been evaluated,
If the desired characteristics are not obtained, a next generation population is generated. As for the selection (step 1), the manner of selection differs depending on the type of the evolution strategy, so here, two representative methods will be described. In the case of an evolution strategy called (μ, λ) -ES, λ generated from μ parent individuals
From the child individuals, μ individuals are deterministically selected in descending order of fitness. In the case of the evolution strategy called (μ + λ) -ES, μ are deterministically selected in descending order of fitness from the population of μ parent individuals and λ child individuals. Evolution strategies include the following methods in addition to the above. When these are used, a selection method is performed in accordance with these methods.・ (1,1) -ES: Random Walk (RW) ・ (1 + 1) -ES: Hill Climbing (HC) ・ (1, λ) -ES, (1 + λ) -ES: Neighborhood Search Method ・ (μ + 1) -ES : Continuous generation type multipoint search method For the crossover (step 2), the normal distribution crossover is used. However, the parent value is inherited for each parameter, and the middle point, inner dividing point, and outer dividing point are used as child values. good. For the mutation (step 3), a perturbation according to a normal distribution is added to each parameter. At this time, the variance of the normal distribution may be adjusted for each parameter, or a correlation between the parameters may be provided. The above-described evolution strategy (ES) has an advantage that conversion from a phenotype to a genotype, such as a genetic algorithm, is not required because each parameter is used as a real value. In addition, by using a crossover method that considers the continuity of real numbers such as normal distribution crossover, it is possible to reduce the traits of the parent rather than one-point or multipoint crossover of binary code or gray code often used in genetic algorithms. It can be strongly reflected in the child's traits.

C. Evolutionary Programming (EP) FIG. 13 is a schematic flowchart of an evolutionary computation module when evolutionary programming is used as an evolutionary computation method. For scaling (step 1), when the number of individuals is μ, q individuals randomly selected from 2 μ individuals each including an individual before the individual perturbation and an individual after the perturbation are applied The number of winners is compared with the fitness of the individual. In the selection (step 2), μ individuals are selected from the generated individuals in order from the one with the highest fitness. The choice is deterministic, but the scaling is stochastic, so the choice is effectively stochastic. In the evolutionary programming (EP) described above, since each parameter is used as a real value, there is an advantage that conversion from a phenotype to a genotype as in a genetic algorithm is not required. Also, since no crossover is used, there is no restriction on the phenotype. The genetic algorithm does not need to make parameters into a string like an evolution strategy, and may use a tree structure or the like.

In the above-described embodiment, an example in which the evolution type calculation is used as the optimization method has been described. However, the optimization method used in the optimization processing unit is not limited to this, and various methods are used. be able to. Hereinafter, some examples of optimization methods other than the evolutionary computation will be described.

1. Neighborhood Search Method Here, as a specific example of the neighborhood search method, an example in which a method combining simulated annealing (SA) and tabu search (TABU) is employed in each optimization processing unit in FIG. 2 will be described. FIG. 14 shows a flow of the entire control when a neighborhood search method combining SA and TABU is used. First, an initial parameter group is generated within a predetermined range (step 1), a test drive (or fuel economy calculation) is performed using the initial parameter group (step 2), and an evaluation value for the result is input (or). Calculation) (step 3). Then, it is determined whether or not the desired drivability and acceleration (or optimal fuel efficiency) are obtained (step 4). If not, the process enters the neighborhood search module (step 5). FIG. 15 is a schematic flowchart showing the processing of the neighborhood search module. First, it is determined whether or not the evaluation of the perturbation solution is higher than the evaluation of the original solution (step 1). If the evaluation of the perturbation solution is higher than the evaluation of the original solution, the solution belongs to the prohibited area (TABU). (Step 2), and if it is not a TABU, add it to the TABU list.
The perturbation solution is moved to the original solution (step 3). Immediately after the start of the optimization process, the original solution does not exist and the TABU list is empty, so the solution of the initial parameter group is set as the original solution, and the TABU list is set. It is added (step 3), and thereafter, it is determined whether or not the temperature is sufficiently low (step 4). In simulated annealing, when the temperature is high, the state transition is also performed with a high probability, and when the temperature is low, a temperature schedule that makes it difficult to perform the state transition is designed in advance,
The search is performed according to the temperature schedule, but the search area of the search is generally global in the initial state and local at the end, that is, the temperature is high in the initial state, and the temperature gradually decreases toward the end. As the temperature schedule is designed. Therefore, since the temperature is high immediately after the start of the optimization process, a perturbation solution is generated without forcibly terminating (step 5), and the temperature parameter is updated according to the temperature schedule (step 6). The perturbation solution is generated by independently applying a perturbation to each component of the current parameter group according to a normal distribution N (0, σ ² ) with mean 0 and variance σ ² . σ is constant, changes adaptively according to the search situation, or is freely set by the user according to the situation. When a perturbation solution is generated, the perturbation solution is used in FIG.
Are performed, that is, test drive (or fuel consumption calculation) (step 2) and evaluation (step 3) are performed. If the desired drivability or optimum fuel consumption is not obtained by the perturbation solution, the neighborhood search module is reentered. (Step 5). The neighborhood search module determines whether the evaluation of the perturbed solution is higher than the evaluation of the original solution (step 1), and if higher than the original solution, determines whether the perturbed solution is a TABU (step 2). ,
If it is not a TABU, it is added to the TABU list, the perturbation solution is set as the original solution (step 4), and it is determined whether or not the temperature is sufficiently low (step 4). Terminate the evolution process and generate a new perturbation solution if the temperature is not low enough. If the evaluation of the perturbation solution is lower than the evaluation of the original solution in the judgment in step 1 described above, the perturbation solution is stochastically set as the original solution according to the temperature (step 6). That is, the higher the temperature, the more globally the search is performed.
It also moves to the perturbed solution side with a lower evaluation than the original solution, but does not move to the perturbed solution side with a lower evaluation than the original solution because the local search is performed when the temperature is low at the end. The above processing
Repeat until the desired drivability and acceleration or optimal fuel economy are obtained, or until the temperature is sufficiently reduced to force the termination. Thus, the local search is gradually performed from the global search to find the optimum value of the parameter. In particular, when this simulated annealing is used for the interactive optimization, the relative evaluation value ΔE indicates how good or bad the evaluation of the current parameter group is based on the evaluation of the previous parameter group. Then, based on this, the movement of the perturbation solution is determined. In the above-described embodiment, a simulated method is used as a neighborhood search method.
Although a method combining annealing and tabu search has been described as an example, it goes without saying that simulated annealing and tabu search may be used alone.

2. Reinforcement learning method Next, an example in which the reinforcement learning method is used as the optimization processing is shown in FIG.
Shown in FIG. 16 is a flowchart schematically showing the processing of the reinforcement learning module. This reinforcement learning module is used in place of the evolutionary calculation module or the neighborhood search module in FIG. 9 or FIG. This method
First, an executable rule is selected for input from the environment. Next, a rule to be executed is determined stochastically (depending on the type of reinforcement learning), and a parameter group is output based on the rule. A reward is given based on the result of activating the parameter group, and the rules are strengthened. In particular, when this is used for interactive optimization, the evaluation by the user is given as a reward. Note that there are two types of reinforcement learning methods: experience-enhancement type and environment identification type. The former emphasizes rewards, and the more successful the rule, the more likely it is to be used. Since the environment identification is emphasized in order to obtain a function that gives a rule to be performed, the rule that is not used has a higher probability of being used.

3. Learning Algorithm + Evolutionary Calculation (or Neighborhood Search Method) Next, FIG. 17 shows an example in which a combination of a learning algorithm and evolutionary calculation is used as optimization processing. FIG.
Is a flowchart schematically showing the processing of the learning + evolution type calculation module. This learning + evolution type calculation module is used in place of the evolution type calculation module or the neighborhood search module in FIG. 9 or FIG. In this method, an evolution type calculation or a neighborhood search method is performed using a degree of connectivity of a neural network constituting a control module or a group of fuzzy rules of a fuzzy system as a parameter group with a throttle opening as input information and throttle characteristics as output information. To optimize the input / output relationship.

4. Evolutionary Computation + Nearby Search Method (Switching Type) Next, FIG. 18 shows an example of using a method in which the evolutionary calculation and the neighborhood search method are switchably combined as optimization processing.
FIG. 18 is a flowchart showing an outline of the processing of the evolutionary calculation + neighborhood search switching type module. This evolutionary calculation + neighborhood search switching type module is the same as the evolutionary calculation module or the neighborhood search module shown in FIG. 9 or FIG. Used instead. According to this method, a global search is performed by an evolutionary calculation, and when a certain degree of convergence is reached, the method is switched to a neighboring step method and a local search is performed, so that efficient optimization can be performed.

[5] Evolutionary Computation + Nearby Search Method (Composite Type) Next, FIG. 19 shows an example in which a method combining the evolutionary calculation and the neighborhood search method in combination is used as optimization processing.
FIG. 19 is a schematic flowchart of the combined learning + evolution type calculation module. This combined learning + evolution type calculation module is used instead of the evolution type calculation module or the neighborhood search module in FIG. 9 or FIG. In this method, a local search is performed on the individual of the evolutionary computation by the neighborhood search method, and the obtained local solution is evolved using the genetic algorithm as an individual, so efficient optimization is possible. . 6. Evolutionary Computation + Taboo Search Finally, an example using a method combining the evolutionary calculation and the tab search as the optimization process will be briefly described. As a result, the evolutionary calculation generates and records the culled individuals in a tab list, and by prohibiting the appearance of the recorded individuals, the same individual is not repeatedly evaluated. Can be reduced.

As in the above-described embodiments, when there are a plurality of control objects and each control module is optimized, an optimization processing unit can be constructed for each control module to be evolved. In addition, the evaluation method of each optimization processing unit in this case is not limited to the combination of the above-described embodiments, and may be any one of the interactive type and the autonomous type, or a combination thereof. In addition, when there are a plurality of control modules and an optimization module is constructed for each of them, depending on the relationship between the control targets, there are some cases in which changes in characteristics affect each other, and optimization of characteristics is a trade-off. It may be. Specifically, for example, the engine and the crane in a crane truck basically do not affect the operating characteristics of both,
Even if the control module that controls the fuel injection device of the engine and the control module that controls the crane are constructed in the same control device, each control module can be optimized without coordination. . However, for example, when the fuel injection device and the electronic throttle device in the same engine are to be controlled, the former control module is optimized to improve the fuel consumption, and the latter control module is optimized to improve the response. However, if the fuel efficiency is improved, the response may be deteriorated, and if the response is improved, the fuel efficiency may be deteriorated. In such a case, optimization processing is coordinated among a plurality of control modules. Specifically, when either the autonomous optimization method or the interactive optimization method is used for the optimization modules of all control modules, the following is obtained after optimizing one control module. By optimizing other control modules to improve or maintain the characteristics, it is possible to optimize the characteristics of other controlled objects within the range of the optimal characteristics of one controlled object, Or by optimizing a plurality of control modules at regular intervals so as to improve or maintain the mutually acquired characteristics, thereby limiting the optimization direction of each control module and performing a plurality of control operations in a short time. To be able to improve the characteristics of the object, or, while optimizing one control module, to improve or maintain the characteristics acquired by that control module. Optimization in parallel to improve the characteristics of one controlled object while obtaining the appropriate characteristics of another controlled object, or By optimizing in parallel so as to improve or maintain the obtained characteristics, it is possible to limit the direction of optimization of each control module and improve the characteristics of a plurality of control targets in a short time. In addition, when using both the autonomous optimization method and the interactive optimization method in combination with the optimization module of a plurality of control modules, the following is required. After optimization, the optimization control module using the autonomous optimization method is used to improve or maintain the obtained characteristics, and the other control modules are optimized to obtain the interactive optimization method. Optimize the characteristics of other controlled objects within the range of the optimum characteristics of the controlled object, or optimize a control module with an optimizing control module using an autonomous optimization method After that, the optimization control module using the interactive optimization method optimizes the other control modules so that the acquired characteristics are improved or maintained. To optimize the characteristics of other controlled objects within the range of the optimum characteristics of the controlled object obtained by the method, or ・ an optimization control module using an interactive optimization method; By repeating the optimization control module using the type optimization method at regular intervals so as to improve or maintain the characteristics obtained from each other, the direction of optimization of each control module is limited and the To be able to improve the characteristics of multiple controlled objects, or ・ Autonomously to improve or maintain the acquired characteristics during the optimization process in the optimization control module using the interactive optimization method The optimization process using the optimization control module using the type optimization method is performed in parallel, so that the characteristics of one control target can be improved by the interactive optimization method and the other control by the autonomous optimization method. Appropriate for the target Optimizing using an interactive optimization method to improve or maintain the obtained characteristics during the optimization process in the optimization control module using the autonomous optimization method By performing optimization processing in parallel in the optimization control module, the characteristics of one control target can be improved by the autonomous optimization method, while appropriate characteristics of the other control target can be obtained by the interactive optimization method Or by optimizing a plurality of control modules in parallel so as to improve or maintain the characteristics acquired by each other, thereby limiting the direction of optimization of each control module and shortening the time. Thus, the characteristics of a plurality of controlled objects can be improved. By coordinating the optimization of a plurality of control modules in the above-described manner, the optimization does not become a trade-off among the control modules. Optimization in a short time. Although it is not possible to cooperate with each other, a plurality of control modules may be independently and parallelly subjected to optimization processing. Sex can be expanded and emergent effects can be expected.

In the above-described embodiment, the basic control module (specifically, for example, the electronic throttle control module in FIG. 2) for determining the operation amount for the control target based on the predetermined input information, or the predetermined input information A correction control module (specifically, for example, a gear ratio correction module or a fuel injection amount correction module in FIG. 2) that determines a correction rate for an operation amount of a control target based on A general characteristic optimization method that generates an initial parameter group, etc., optimizes control parameters using an optimization method, and updates the control parameters of the basic control module or the correction control module to optimized control parameters one after another. As described above, the method for optimizing the overall characteristics is not limited to the above-described embodiment, and the normal control method may be used. Any method may be used as long as it optimizes the control module for use, for example, a. When the control module for normal control is a control module for correction that outputs a correction amount with respect to the output of the basic control module, the control module for optimization having the same control parameters as the control module for normal control (correction control module) After optimizing the control parameters of the optimization control module, the control parameters of the normal control module (correction control module) may be updated to the optimized control parameters (FIG. 20).
B). When the control module for normal control is a control module for correction that outputs a correction amount to the output of the basic control module, and further includes a module for learning and a module for execution, the control module for normal control (for correction) A control module for optimization having the same control parameters or input / output relation as the control module) is provided, and after optimizing the control parameters or input / output relation in the control module for optimization, the control module for normal control (for correction) The learning module in the control module) may learn the optimized control parameters or the input / output relationship, and after the learning of the learning module is completed, the learning module and the execution module may be exchanged (FIG. 21). C.), C. An initial value of a control parameter of the control module for normal control may be determined in advance, and a correction amount or a correction rate of the initial value may be optimized using an optimization method (see FIG. 22), and d. When a linear function is used as the control module for normal control, a control module for optimization configured to output control parameters of the control module for normal control based on predetermined input information is provided. A method of optimizing the control parameters by optimizing the control module may be used (see FIG. 23). In the case of the above method d, the optimization module may be constructed in any manner. For example, the optimization module outputs the control parameter value of the normal control control module based on predetermined input information. When the neural network is constructed, the coupling coefficient or its input / output relationship can be optimized. When the optimization module is constructed by fuzzy inference, the rule is optimized. Can be The control module for normal control in the methods c and d may be a basic control module or a correction control module.

[0023]

As described above, according to the first aspect of the present invention,
The optimization coordination method in the characteristic optimization method according to the first aspect of the present invention is a characteristic optimization method for optimizing characteristics of a plurality of normal control modules for determining an output related to an operation amount of a control target based on predetermined input information. In the method, after optimizing one control module for normal control, optimization of another control module for normal control is performed so as to improve or maintain the obtained characteristics. This has the effect that the characteristics of other control objects can be optimized within the control. The optimization coordination method in the characteristic optimization method according to claim 2 of the present invention is characterized in that each of the characteristics of a plurality of normal control control modules for determining an output related to an operation amount of a control target based on predetermined input information is determined. In the characteristic optimization method for optimizing, a plurality of normal control control modules are optimized at regular intervals so as to improve or maintain the mutually acquired characteristics. Therefore, the optimization direction of each normal control control module is optimized. There is an effect that the characteristics of a plurality of controlled objects can be improved in a short time by limiting the performance. The optimization coordination method in the characteristic optimization method according to claim 3 of the present invention is characterized in that each of the characteristics of a plurality of normal control control modules for determining an output related to an operation amount of a control target based on predetermined input information is determined. In the optimization method for optimizing characteristics, while optimizing one control module, optimization of other control modules is performed in parallel so as to improve or maintain the characteristics obtained by the control module. There is an effect that appropriate characteristics of another control target can be obtained while improving characteristics of the control target. The optimization coordination method in the characteristic optimization method according to claim 4 of the present invention is characterized in that each of the characteristics of a plurality of control modules for normal control that determines an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimizing method to be optimized, a plurality of normal control control modules are optimized in parallel so as to improve or maintain the characteristics obtained by each other. And the characteristics of a plurality of controlled objects can be improved in a short time. An optimization coordination method in the characteristic optimization method according to claim 5 of the present invention is the coordination method according to any one of claims 1 to 5, wherein at least one normal control control module is optimized. Using an autonomous evaluation method that performs an evaluation during the optimization process based on a preset evaluation criterion, optimization of the other control modules for normal control is performed based on the evaluation based on the user's will. Since an interactive evaluation method for performing an evaluation is used, for example, after optimizing one control module for normal control using an interactive optimization method, another control module for normal control is used to improve or maintain the obtained characteristics. When optimizing a control module, it is possible to optimize the characteristics of other controlled objects within the range of the optimum characteristics of the controlled object obtained by the interactive optimization method. Evaluation method After optimizing one normal control module by using it, when optimizing another normal control module to improve or maintain the obtained characteristics, the autonomous optimization method was used. It is possible to optimize the characteristics of other controlled objects within the range of the optimum characteristics of the controlled object, and to optimize the control module for normal control using an interactive evaluation method and autonomous optimization When the optimization of the control module for normal control using the method is repeated at regular intervals so as to improve or maintain the characteristics obtained from each other, the direction of optimization of each control module is limited to shorten it. The characteristics of multiple controlled objects can be improved in time, and the characteristics obtained during the optimization process of one control module for normal control using the interactive evaluation method When performing optimization processing of other control modules for normal control in parallel using the autonomous optimization method so as to improve or maintain, while improving the characteristics of one controlled object by the interactive optimization method In addition, appropriate characteristics of other controlled objects can be obtained by the autonomous optimization method, and while the optimization process of one control module for normal control using the autonomous evaluation method is performed, If the optimization processing of other control modules for normal control is performed in parallel using the interactive optimization method so as to improve or maintain the characteristics that have been performed, the characteristics of one control target can be adjusted by the autonomous optimization method. While improving, it becomes possible to obtain appropriate characteristics of other controlled objects by the interactive optimization method. If necessary, for example, for one control target, a reference characteristic of the control target with time-dependent deterioration or an evaluation reference based on the regulation of the control target is set in advance and used within the range of the evaluation reference. If constructed so that evaluation based on the user's will can be performed, there is an effect that the characteristics can be optimized according to the user's preference within the range of regulation.

[Brief description of the drawings]

FIG. 1 is a schematic diagram showing a relationship between an engine 1 and a control device 10 that executes a comprehensive characteristic optimizing method.

FIG. 2 is a schematic block diagram of a control device 10.

FIG. 3 is a schematic block diagram of an electronic throttle control module.

FIG. 4 is a graph showing an example of static characteristics of some throttles.

FIG. 5 is a graph showing examples of dynamic characteristics of some throttles.

FIG. 6 is a diagram illustrating a coding example of a control parameter in the electronic throttle control module.

FIG. 7 is a schematic block diagram of a fuel injection device control module.

FIG. 8A is a schematic diagram of a neural network constituting a fuel injection amount correction module, and FIG. 8B is a diagram illustrating a coding example of control parameters of a fuel injection control module.

FIG. 9 is a flowchart showing a flow of an optimization process of the entire control device 10;

FIG. 10 is a diagram illustrating an example of a dividing method when each individual is evaluated in a time-sharing manner.

FIG. 11 is a schematic flowchart of an evolutionary calculation module when a genetic algorithm is used as the evolutionary calculation method.

FIG. 12 is a schematic flowchart of an evolutionary computation module when an evolutionary strategy is used as an evolutionary computational method.

FIG. 13 is a schematic flowchart of an evolutionary computation module when evolutionary programming is used as the evolutionary computation method.

FIG. 14 shows an overall control flow when a neighborhood search method combining simulated annealing and tabu search is used.

FIG. 15 is a schematic flowchart showing processing of a neighborhood search module.

FIG. 16 is a flowchart schematically showing processing of a reinforcement learning module.

FIG. 17 is a flowchart schematically showing processing of a learning + evolution type calculation module.

FIG. 18 is a flowchart showing an outline of processing of an evolutionary calculation + neighborhood search switching type module.

FIG. 19 is a schematic flowchart of a learning + evolution type calculation combined module.

FIG. 20 is a diagram showing another embodiment of the comprehensive characteristic optimizing method.

FIG. 21 is a diagram showing still another embodiment of the comprehensive characteristic optimizing method.

FIG. 22 is a diagram showing still another embodiment of the comprehensive characteristic optimizing method.

FIG. 23 is a diagram showing still another embodiment of the comprehensive characteristic optimizing method.

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) F02D 45/00 340 F02D 45/00 340Z 374 374Z F-term (Reference) 3D041 AA21 AA31 AA32 AA66 AB01 AB07 AC19 AD00 AD02 AD05 AD10 AD14 AD48 AE04 AE07 AE36 AF05 AF07 AF09 3G084 AA06 BA05 BA13 DA02 DA05 EB08 EB17 EB20 FA00 FA01 FA02 FA10 FA11 FA20 FA33 3G093 AA01 AA10 BA14 BA16 BA19 BA27 DA03 DA04 DA05 DA06 DB08 DB21 EA05 FA02 FA02 FA02 JA03 LA01 LB01 LC01 LC03 MA01 MA11 NC02 ND18 ND25 ND43 PA07Z PA09Z PA10Z PA11A PA11Z PB03A PB03Z PE01Z PE08Z PF08A PG00Z 5H004 GA21 GA26 GB12 HA01 HA03 HB02 JA03 KD62 KD63 KD67 KD67 MA12

Claims (14)

[Claims] 1. A characteristic optimizing method for optimizing characteristics of a plurality of normal control modules for determining an output related to an operation amount of a control target based on predetermined input information. An optimization coordination method in the characteristic optimization method, characterized in that after optimizing the module, another normal control module is optimized so as to improve or maintain the acquired characteristics. 2. A characteristic optimization method for optimizing characteristics of a plurality of control modules for normal control, each of which determines an output related to an operation amount of a control target based on predetermined input information, comprising: An optimization coordination method in a characteristic optimization method characterized by performing optimization at regular intervals so as to improve or maintain mutually acquired characteristics. 3. A characteristic optimizing method for optimizing characteristics of a plurality of control modules for normal control, each of which determines an output related to an operation amount of a control target based on predetermined input information. In the characteristic optimization method, optimization of another normal control module is performed in parallel so as to improve or maintain the characteristic obtained by the normal control module during optimization of the module. Optimization coordination method. 4. A characteristic optimizing method for optimizing characteristics of a plurality of normal control modules for determining an output related to an operation amount of a control target based on predetermined input information. An optimization coordination method in a characteristic optimization method, wherein modules are optimized in parallel so as to improve or maintain mutually acquired characteristics. 5. An autonomous evaluation method for performing an evaluation during an optimization process based on a preset evaluation criterion is used for optimizing at least one normal control control module. The characteristic optimization method according to any one of claims 1 to 5, wherein the optimization is performed using an interactive evaluation method that performs an evaluation during the optimization process based on an evaluation based on a user's intention. Coordination method of optimization in Japan. 6. The comprehensive characteristic optimizing method according to claim 5, wherein said evaluation criterion is set based on a reference characteristic of a control object accompanied by deterioration with time. 7. The comprehensive characteristic optimization method according to claim 5, wherein said evaluation criterion is set based on regulation for a control object. 8. A characteristic optimizing method for optimizing each of a plurality of characteristics of at least one normal control control module for determining an output related to an operation amount of a control target based on predetermined input information, wherein: Optimizing, and then optimizing other characteristics so as to improve or maintain the obtained characteristics. 9. A characteristic optimization method for optimizing each of a plurality of characteristics of a control module for normal control for determining an output related to an operation amount of a control target based on predetermined input information, A optimization coordination method in the characteristic optimization method, wherein the optimization is performed at regular intervals so as to improve or maintain the characteristics acquired at the same time. 10. A characteristic optimizing method for optimizing a plurality of characteristics of a control module for normal control which determines an output related to an operation amount of a control target based on predetermined input information, wherein one characteristic is optimized. A method of coordinating optimization in a characteristic optimization method, wherein other characteristics are optimized in parallel so as to improve or maintain the characteristics. 11. A characteristic optimizing method for optimizing each of a plurality of characteristics of a control module for normal control for determining an output related to an operation amount of a control target based on predetermined input information, Optimization in a characteristic optimization method characterized by performing parallel optimization so as to improve or maintain the characteristics obtained by the optimization. 12. An autonomous evaluation method for performing an evaluation during an optimization process based on a preset evaluation criterion is used for optimizing at least one normal control control module. The characteristic optimization method according to any one of claims 8 to 11, wherein the optimization is performed using an interactive evaluation method that performs an evaluation during the optimization process based on an evaluation based on a user's will. Coordination method of optimization in Japan. 13. The comprehensive characteristic optimizing method according to claim 12, wherein said evaluation criterion is set based on a reference characteristic of a control object accompanied by deterioration with time. 14. The comprehensive characteristic optimizing method according to claim 12, wherein said evaluation criterion is set based on regulation for a control object.