JP4850754B2 - Automatic adjustment system for control parameters - Google Patents

Description

  The present invention relates to a control parameter automatic adaptation system for adapting a control parameter related to control of a moving body to an optimum value.

  2. Description of the Related Art Conventionally, in order to automatically obtain (automatically adjust) an optimum value of a control parameter of a moving body such as an automobile, a technique for learning an optimum parameter based on an evaluation value obtained by simulation is known.

  However, in the method of obtaining an evaluation value by simulation, there is a problem that a parameter acquired by learning is not an effective parameter when applied to an actual machine due to an approximation error of the simulation model.

  For this reason, recently, an attempt has been made to introduce a real machine into a learning loop and obtain an evaluation value directly from the real machine. For example, Patent Document 1 proposes a technique for learning using an actual engine in a form that replaces a simulation model and acquiring a control parameter that matches a target evaluation condition.

Patent Document 2 discloses a technique for performing a learning operation in a computer based on an evaluation value set sampled in advance from real machines of an engine or a motor as a method for learningly improving evaluation values of a plurality of objective functions. ing.
JP 2004-116351 A JP 2005-285090 A

  However, since it takes a long time to obtain an evaluation value in an actual machine compared to simulation, simply changing the acquisition method of the evaluation value from simulation to actual machine without improving the learning logic is optimal in realistic time. It is difficult to obtain parameters.

  Also, unlike simulations, actual machines have problems such as measurement noise, variation in control results, drift in characteristics of actual machines such as vehicles and engines, etc., and it is difficult to treat evaluation values as deterministic. Optimal parameters cannot be obtained by simply improving the learning progress without improving the value handling.

  In this regard, although the technique of Patent Document 1 addresses the above-described problem by performing a step-by-step evaluation of individuals using a genetic algorithm, there are a plurality of degrees of fitness when evaluating superiority or inferiority between individuals. Since the linear sum of the evaluation values is used, the designer must determine the weighting coefficient for determining the balance between the evaluation values by trial and error, and there is room for improvement in terms of efficiency.

  In Patent Document 2, first, representative sample points are acquired from an actual machine, and later learning calculation is performed using estimated evaluation values calculated from the sample points. For this reason, although learning time can be shortened in that the evaluation value is not obtained directly from the actual machine during learning, it is necessary to obtain a sample in advance. Since the value at is neglected by treating it deterministically, it is insufficient to solve the problem.

  Further, in Patent Document 2, it is conceivable that the estimated evaluation value is different from the true evaluation value, and the obtained control parameter is effective when applied to an actual machine as in the case of the model approximation error which is a problem in obtaining the evaluation value by simulation. The case which is not cannot be excluded.

  The present invention has been made in view of the above circumstances, and while eliminating the approximation error of the simulation model using a real machine, it suppresses the influence of increase in learning time due to the use of the real machine and fluctuations in characteristics and drift, and is a good optimal An object of the present invention is to provide a control parameter automatic adaptation system capable of obtaining parameters quickly and efficiently.

In order to achieve the above object, a control parameter automatic adaptation system according to the present invention is an automatic control parameter adaptation system that adapts a control parameter related to control of a moving body to an optimum value, and learns a real machine as the mobile body. Introduced into a loop, equipped with a learning operation unit that automatically obtains evaluation values by automatic operation of the actual machine, independently evaluates them with multiple evaluation axes, and obtains control parameters that are optimal solutions by a multi-purpose genetic algorithm The learning operation unit calculates a fitness of an individual including the control parameter as a gene, and when selecting a part of individuals from a plurality of individuals as a parent individual. A parent selection unit for determining superiority or inferiority of solutions using an index according to the degree of density with other individuals when the fitness of the individuals is equivalent; and the selected parent individuals A generation alternation unit that generates a child individual by genetic manipulation, a child individual evaluation unit that selectively evaluates the generated child individual multiple times, and a survival selection unit that selects an individual to be stored in the next generation Yes, and said child individual evaluation unit only if the generated child individual is determined to Pareto optimal solutions in the evaluation for the first time, and wherein the row Ukoto the re-evaluation of the child individuals.

  According to the present invention, while eliminating the approximation error of the simulation model using the real machine, it is possible to suppress the influence of the increase in the learning time and the characteristic variation / drift due to the use of the real machine. Can be obtained efficiently.

  Embodiments of the present invention will be described below with reference to the drawings. 1 to 11 relate to an embodiment of the present invention, FIG. 1 is an overall configuration diagram of an automatic adaptation system, FIG. 2 is an explanatory diagram showing a speed pattern, FIG. 3 is a block diagram of an automatic operation control system, and FIG. Is a block diagram of a learning operation unit, FIG. 5 is an explanatory diagram showing an example of a solution group, FIG. 6 is an explanatory diagram of a niche size, FIG. 7 is an explanatory diagram showing an example of calculating a niche count, and FIG. 8 is a flowchart of an optimization process FIG. 9 is a flowchart of the child individual evaluation process, FIG. 10 is an explanatory diagram showing an example of the result of the child individual reevaluation, and FIG. 11 is an explanatory diagram showing the optimization result.

  The automatic adjustment system for control parameters according to the present invention optimizes parameters related to the control of a moving body operated by a person such as an automobile, and is configured as an automatic adjustment system 1 shown in FIG. The automatic adaptation system 1 obtains a target evaluation value by directly operating an actual machine without using a simulation, and automatically adjusts the optimum control parameter by learning.

  Specifically, the automatic adaptation system 1 controls the actual machine 10 (for example, a power unit system such as an automobile engine or motor, a transmission system, a brake system, or an actual vehicle) whose control parameters are to be optimized. An automatic driving device 20 that automatically operates the actual machine 10 as a controller that performs the measurement, a measuring instrument 30 that includes various sensors that measure various data under the automatic operating state of the actual machine 10, and data obtained through the measuring instrument 30 An evaluation value calculation unit 40 that calculates a target evaluation value, a learning calculation unit 50 that updates and optimizes a control parameter to a better value based on the calculated evaluation value, and an updated parameter ( A parameter update unit 60 is provided to the automatic driving device 20).

  The optimization of the control parameters by the learning calculation unit 50 is performed using a multi-objective optimization method, and the superiority or inferiority of the control parameters is determined on the evaluation space having each target evaluation item as an axis, and control is performed based on the superiority / inferiority determination. The parameters are updated using a genetic algorithm. Then, this control parameter is given to the actual machine for automatic operation, and learning is repeated by repeatedly evaluating the operation result in the evaluation space.

  In the following, as shown in Fig. 2, there are two speed error RMS (Root Mean Square) [km / h] and fuel consumption rate [km / L] when the car is driven along a specified speed pattern. An example of selecting an evaluation function for multi-objective optimization and optimizing the accelerator and brake operation gain as design variables will be described.

  In this case, the fuel consumption rate is expected to increase when the speed is reduced on the acceleration side and the fuel consumption is reduced, while the speed is kept high on the deceleration side and the mileage is increased, while the speed error from the speed pattern is Although a trade-off relationship of increasing is assumed, by using a multi-purpose genetic algorithm, a design variable set that maintains a good balance of the trade-off relationship can be obtained.

  Specifically, as shown in FIG. 3, the automatic driving device 20 that controls the follow-up operation to the specified speed pattern uses the automobile 11 on the chassis dynamometer (CDM) as an actual machine 10 and performs speed control by an FA computer or the like. And a driver 22 that receives the output from the speed controller 21 and drives the accelerator and brake of the automobile 11.

  The speed control of the automobile 11 via the driver 22 includes an electric accelerator operation in which the output voltage of an accelerator position sensor (APS) that detects the depression position of the accelerator pedal is varied and output to the electronically controlled throttle device of the automobile 11. This is performed by a mechanical brake operation in which the brake pedal is depressed by an actuator (for example, an electric slider mechanism) attached in the vicinity of the brake pedal of the automobile 11.

  The speed controller 21 calculates accelerator and brake operation amounts in accordance with a speed control law based on a preset driver model, converts these operation amounts into APS voltage and brake pedal stroke, and operates the accelerator and brake. And a gain changing unit 21b that changes the gain and makes the vehicle characteristics variable. In this embodiment, since the speed control law (driver model) is fixed, changing the operation gain of the accelerator and the brake may cause the vehicle side response to the operation intention to be slow or sensitive when viewed from the driver model side. It corresponds to being changed.

  Driving data of the automobile 11 that is automatically driven by the automatic driving device 20 is sent to the evaluation value calculation unit 40 via the measuring device 30. The evaluation value calculation unit 40 calculates the fuel consumption rate and the RMS of the speed error with respect to the travel pattern as the evaluation value from the operation data acquired via the measuring device 30, and sends the evaluation value to the learning calculation unit 50.

  The evaluation value calculation unit 40, the learning calculation unit 50, and the parameter update unit 60 may be configured as a part of the functions of the FA computer that configures the automatic driving device 20. It may be configured by a single computer or a plurality of computer devices.

  The learning calculation unit 50 optimizes the accelerator and brake gains (scalar values) as genes of the multipurpose genetic algorithm, and adaptively learns and acquires design variables with the best balance of evaluation values. In this embodiment, the gene is handled by normalizing to a range of 0 to 1, and this normalized range corresponds to a gain of 0.3 to 1.5. The gain 1.0 is a design value of the automatic driving device 20.

  The optimization in this form is based on the SPEA2 (Strength Pareto Evolutionary Algorithm 2) method, which is a kind of multi-objective optimization method. However, the following (1) As shown in (3), it is necessary to consider the uncertainty of the evaluation value, the limitation on the number of trials, and the drift of vehicle characteristics.

  Regarding SPEA2, for example, E. Zitzler, M. Laumanns, L. Thiele: SPEA2: Improving the Performance of the Strength Pareto Evolutionary Algorithm, Technical Report 103, Computer Engineering and Communication Networks Lab (TIK), Swiss Federal Institute of Technology ( ETH) Zurich (2001) and the like.

(1) Uncertainty of evaluation values Even if the same design variables (constants) are given, vehicle behavior varies with a certain distribution. For this reason, it is desirable to perform evaluation multiple times to obtain the distribution and the average value. However, there is a problem of limitation of the number of trials shown in the following (2), which requires some ingenuity.

(2) Limiting the number of trials The next problem is the limitation on the number of evaluations. In particular, in order to obtain an evaluation value, it is necessary to travel a speed pattern of a certain length of time, while in order to obtain an effective solution within a realistic time, it is necessary to reduce the number of evaluations. Therefore, if the number of individuals in the multipurpose genetic algorithm is reduced, there is a possibility that the shape of the trade-off boundary surface cannot be grasped sufficiently with a small number of individuals.

(3) Drift of vehicle characteristics Even if a vehicle is warmed up on the CDM, the vehicle characteristics will gradually change due to changes in water temperature, oil temperature, tire pressure, and other characteristics as optimization is performed. Go. If this characteristic drift is not taken into account, there is a possibility that the evaluation value evaluated at the initial stage of optimization loses reliability at the end of optimization.

  The problems (1) to (3) described above are achieved by introducing the methods of “re-evaluation of child individuals”, “sharing evaluation”, and “priority evaluation of child individuals”, respectively, as described below. Can be solved. For this reason, as shown in FIG. 4, the learning calculation unit 50 includes an initialization unit 51, a fitness calculation unit 52, a parent selection unit 53, a generation change unit 54, a child individual evaluation unit 55, a survival selection unit 56, a life determination. Unit 57 and end determination unit 58. The parent selection unit 53, the child individual evaluation unit 55, and the survival selection unit 56 respectively include "sharing evaluation", "re-evaluation of child individual", and "child individual evaluation". Priority evaluation ”is implemented.

  The initialization unit 51 uses control parameters to be optimized as genes, and randomly generates individuals having these genes according to a random or predetermined rule. In this embodiment, an accelerator gain and a brake gain are used as genes, and an individual having these genes is generated as an initial individual. The number of initial individuals corresponds to the limitation on the number of trials described above, so the number of individuals is smaller than the number of individuals by normal simulation, for example, the number of individuals is reduced from 100 to 10 and 9 out of 10 individuals are designed variables. Are arranged in a grid pattern in the value range, and the remaining one individual is arranged as a random value.

  The fitness level calculation unit 52 calculates the fitness level for each individual. In SPEA2 used in this embodiment, evaluation values f1k, f2k,... For a plurality of evaluation criteria are individually assigned to one individual Ik, and the dominance relationship (dominance relationship) between individuals based on these evaluation values. ), The fitness Qk of each individual Ik is calculated.

  Specifically, the fitness calculation unit 52 selects one individual Ik from each individual, sends the control parameters P1k to Pnk of the individual Ik to the automatic driving device 20, performs the evaluation operation, and measures the measuring device 30. Evaluation values f1k, f2k,... For a plurality of evaluation criteria are calculated for each individual Ik based on the response Rk such as the control result output from the control result output from the automatic driving device 20.

  In this embodiment, the evaluation criteria are the fuel consumption rate and the speed error (RMS), the evaluation value f1 is the evaluation value for the fuel consumption rate, the evaluation value f2 is the evaluation value for the speed error, the evaluation value f1 is the horizontal axis, and the evaluation value When each individual Ik is plotted on the evaluation plane with f2 as the vertical axis, for example, a solution population as shown in FIG. 5 is obtained. In this embodiment, the larger the evaluation value f1 and the smaller the evaluation value f2, the better the individual.

  In calculating the fitness Qk, first, for all individuals Ik, the number of other individuals that dominate the individual Ik (the number of superior individuals; the number of dominant individuals) is obtained, and then, for each individual, Add all the control numbers of other individuals who control the individual. A value obtained by adding the control numbers is the fitness Qk of each individual Ik. When an individual is not controlled by any individual, the fitness is 0, and the individual is a Pareto optimal solution (hereinafter simply referred to as “Pareto solution”). In the example of FIG. 5, an individual indicated by a circle is a Pareto solution, and a broken line connecting these Pareto solutions forms a boundary (a Pareto boundary) for a non-Pareto solution (inferior solution) indicated by a square mark.

  The parent selection unit 53 selects a predetermined number of individuals (for example, three individuals) smaller than the total number of individual groups from the individual groups as parent individuals. In the selection of the parent individual, when the child individual generated in the previous generation is a Pareto solution, the individual is selected as one of the parent individuals in the current generation, and the child individual generated in the previous generation is selected. Is not a Pareto solution, it is selected by a selection method such as a tournament method.

  In this case, if individuals with similar fitness are selected as parent candidates, an index (density niche count) of how close other individuals are around is calculated and In the sharing, the individuals with lower density are preferentially used as parent individuals. That is, when the number of individuals is small (in this embodiment, 10 individuals as initial individuals), a plurality of solutions are collected in one place in the evaluation space, and it is difficult to determine the shape of the Pareto boundary surface There is. In such a case, generally the distributed logic of the solution by the “end-cut method” is adopted, but the “end-cut method” is used when all the solutions including the child individuals are determined as Pareto solutions. This algorithm is only activated and does not affect the initial learning solution.

Therefore, in this embodiment, in the case of selecting the parent of the tournament method of the genetic algorithm, an evaluation method by sharing is introduced in addition to the SPEA2 fitness index. Specifically, when the fitness is the same value, a niche count based on the sharing function s (d) shown in the following formula (1) is calculated and applied to the fitness, and an individual with a lower density (niche count) Is the parent individual.
s (d) = 1- (d / ds) ^α d <ds (1)
= 0 d ≧ ds

  Here, as shown in FIG. 6, d is the Euclidean distance between the individual i and the individual j in the objective function space, and ds is how close the position is in the similar individuals when the evaluation is lowered. Is the maximum distance for estimating, and is called the niche size. The niche size ds is defined by a value specified in advance, a ratio to the variance of the evaluation value or the control parameter value (for example, about 30% of the control parameter value range). Α is a power coefficient.

The niche count n _i of the individual i can be obtained as the sum of the sharing functions s _j (d) of the individual j existing within the niche size ds, as shown in the following equation (2). The greater the niche count n _i is increased, will be are densely individuals, it becomes small fitness of the individual i, can be determined Pareto solutions with more expansive. FIG. 7 shows an example of calculation of the niche count, and the control number, fitness, and niche count values of the individual are added to the upper, middle, and lower stages of the circles representing the individual, respectively.
n _i = Σs _j (d) (2)

  With the above sharing strategy, the selectivity of the parent individuals in the area where the solutions are dense can be lowered, and as a result, the child individuals can be dispersed. This makes it possible to obtain a solution set that is appropriately distributed in diversity and solution space with a small number of individuals, and this algorithm also works for parent selection at the time of crossover described later. It can act on the evolution strategy and disperse the solution at an early stage.

  The niche count can be defined on both the evaluation plane and the design variable plane (parameter plane), but it is preferable to evaluate the parameter plane from the viewpoint of diversity with respect to the control parameters after optimization.

  The generation change unit 54 generates a child individual from the parent individual selected by the parent selection unit 53 by genetic operation. In this embodiment, one child individual is generated in each generation by a crossover operation from three parent individuals. This is because the latest child individual is always reflected in the archive group that is the target of parent selection.

  As the crossover operation, for example, UNDX (Unimodal Normal Distribution Crossover) is used. In UNDX, a child individual is generated according to a normally distributed random number determined based on the positional relationship between three parent individuals. For example, assuming that there are two control parameters to be optimized, and if the first and second parent individuals are arranged on the parameter plane, the child individuals generate normal random numbers according to the normal distribution around the axis connecting the two parent individuals. Generated using. The third parent individual is used as an auxiliary to determine the component of the axial standard deviation orthogonal to the axis connecting the two parent individuals, and the child individual is far away from the first to third parent individuals. It is not generated at the location.

  Regarding UNDX (unimodal normal distribution crossover), see, for example, edited by Hiroaki Kitano, “Genetic Algorithm 4”, Sangyo Tosho Co., Ltd., p.232-235.

  The child individual evaluation unit 55 performs actual machine evaluation using the generated control parameters of the child individual. If it is determined as a Pareto solution as a result of the actual machine evaluation, the actual machine is re-evaluated at least once with the control parameter of the child individual. This is to suppress an increase in the number of evaluations, such as evaluating all individuals multiple times, by re-evaluating only when the child individual is determined to be a Pareto solution at the time of the first determination, It is possible to reduce the time related to the evaluation value acquisition by the actual machine.

  In the re-evaluation of the child individual, the determination is made again on the evaluation space based on the average and variance of the plurality of evaluation values, and the Pareto solution is first survived as the Pareto solution when it is determined again. In addition, a child individual determined to be a non-Pareto solution (inferior solution) at the time of re-evaluation is discarded in the first Pareto determination and treated in the same manner as a solution determined to be inferior at the time of initial determination.

  That is, by setting the evaluation value obtained by re-evaluation as a new evaluation value, a stable Pareto solution can be obtained by eliminating variations in actual machine characteristics. As a result of re-evaluation, if the difference between the evaluation values is equal to or greater than the prescribed value, or if variance of the evaluation values exceeding the variation of the evaluation values predicted in advance is detected, a part of the evaluation values Alternatively, the abnormal value can be detected and eliminated by discarding the child individual itself as an abnormal value and performing reevaluation.

  The survival selection unit 56 performs the travel evaluation using the generated control parameter of the child individual. As a result, when the matching degree is the same as that of the existing parent individual, the child individual is given priority in the upper selection. This is because, as described above, when the evaluation value is calculated in an actual vehicle, the reliability of the evaluation value may decrease due to the change in the vehicle characteristics when the time elapsed after obtaining the evaluation, By performing the survival selection with priority on the individual child, priority is given to the more recent evaluation, and the difference between the evaluation value and the actual characteristic can be suppressed.

  In addition, the survival selection unit 56 stores the number of generations that each individual has been generated, holds the age of each individual, and when it becomes an equivalent fitness at the time of selection for survival in the next generation, Spear older individuals.

  This individual age can also be applied as one of the above-mentioned parent individual selection conditions. If the fitness is the same, a younger parent may be selected preferentially.

  The life determination unit 57 determines whether or not the age of each individual has reached a predetermined number of generations (life), and when the life is exceeded, re-evaluates with the control parameters of the individual, Check gender regularly. As a result, the Pareto solution may be changed to a non-Pareto solution, and vice versa.

  The end determination unit 58 determines whether or not the optimization end condition is satisfied for the next-generation individual selected in this manner, and repeats the above evolution process unless the end condition is satisfied. When it is determined that the end condition is satisfied, the control parameter of the individual that is a Pareto solution is output as an optimized control parameter. The optimization termination condition is, for example, a preset time or the number of generations.

  Next, the optimization process in the above automatic adaptation system 1 is demonstrated using the flowchart of FIG.

  FIG. 8 is a flowchart showing the overall flow of the optimization process in the automatic adaptation system 1. In the first step S1, as an initialization process of the evolution process of the genetic algorithm, first, the initialization unit 51 performs a predetermined number of processes. Create an initial individual. Next, in step S2, the initial individual control parameters are given to the automatic driving device 20 to perform the initial evaluation operation by the driving operation on the CDM of the actual machine (the automobile 11), and the fitness is calculated for each initial individual. Try to do. In the second and subsequent generations, only one generated child individual is tried, and the number of trials per generation is reduced.

  After the initial evaluation operation in step S2, the process proceeds to step S3, the evaluation value of each individual is calculated from the data acquired via the measuring device 30, and the fitness of each individual is calculated based on this evaluation. . After calculating the fitness of the initial individual, generational change is performed in step S4 and subsequent steps to select a parent individual and generate a child individual.

  In this generation change, an individual with a low fitness is selected as a parent individual, and if the fitness is the same, an individual with a low density is preferentially selected as a parent by sharing. That is, in step S4, the density by the above-described niche count is calculated for each individual, and in step S5, an individual having a low fitness and density is selected as a parent individual, and in step S6, the selected parent individual is crossovered. Etc. to generate child individuals. In this embodiment, only one individual child is generated in one generation.

  Thereafter, in step S7, the child individual generated by the generation change is evaluated by the actual machine to determine whether or not it is a Pareto solution. In this Pareto solution determination, even if it is determined to be a Pareto solution in the initial determination, the determination is not confirmed immediately, but by performing multiple re-evaluations, it is overestimated due to variations in actual machine characteristics, etc. Eliminate individuals that are Pareto solutions.

  In this embodiment, the evaluation of the child individual is performed in the child individual evaluation sub-process shown in the flowchart of FIG. 9, and the determination of the Pareto solution is finalized after two evaluations. In this sub-process, in the first step S21, the actual machine running operation is performed with the control parameters of the child individuals, and in step S22, the evaluation value is calculated from the actual machine running operation result. Next, in step S23, the fitness of the individual is calculated from the evaluation value in the actual machine driving operation, and in step S24, it is determined whether or not the Pareto solution.

  If the result of determination in step S24 is that the generated child individual is not a Pareto solution, this sub-process is exited, and if it is determined that the generated child individual is a Pareto solution, the process proceeds from step S24 to step S25. In step S25, an actual machine running operation for re-evaluation is performed, and the average evaluation value is calculated by averaging two evaluation values in step S26. In step S27, the fitness is calculated based on the average evaluation value. In step S28, it is determined again whether the pareto solution is used.

  As a result, if it is determined that it is a Pareto solution in the second determination in step S28, it is determined in step S29 that the child individual is a Pareto solution and this sub-process is exited, and it is determined that the child is not a Pareto solution If so, the first pareto determination is discarded in step S30, and the child individual is regarded as a non-pareto solution (inferior solution) and the present sub-process is exited.

  For example, in the example shown in FIG. 10, the total number of offspring individuals is 77 and 20 individuals were determined to be Pareto solutions in the initial evaluation, whereas the average evaluation value after re-evaluation is half 10 individuals are determined to be inferior (non-Pareto solution) or cocoon (individual not remaining in the next generation). Therefore, it can be said that the Pareto solution with high reliability remains in the next generation due to the effect of the reevaluation of the offspring.

  Next, after evaluating the child individual, the process proceeds to step S8 in the overall optimization process of FIG. 8, and the survival selection process is performed. In this survival selection process, in order to follow the drift of the vehicle characteristics and acquire more effective control parameters in the actual machine, if the newly generated child individual and the existing parent individual have the same fitness, The individual is prioritized to survive to the next generation, and the age of each individual is compared.

  Each individual selected in step S8 is determined in step S9 whether or not its age has reached a predetermined number of generations (lifetime). If it is determined that the lifetime has been exceeded, in step S10, after performing the reevaluation operation with the control parameter of the individual, the process returns to step S3 and the above processing is repeated to confirm the effectiveness of the control parameter. To do.

  If it is determined in step S9 that the lifetime has not been reached, it is determined in step S11 whether or not an optimization termination condition (preset time, number of generations, etc.) has been satisfied. If not satisfied, the process returns to step S3 through the re-evaluation operation in step S10, and the above evolution process is repeated. When it is determined that the termination condition is satisfied, the optimization process is terminated, and the control parameter of the individual that is the Pareto solution is output as the optimized control parameter.

  FIG. 11 shows an example of the evolution result by the above optimization processing on the parameter plane. In FIG. 11, individuals indicated by triangles are initial values before optimization, and individuals indicated by black circles are optimized. Individuals indicated by the later inferior solution and hatched circles indicate the Pareto solution after optimization. From FIG. 11, it can be seen that there is a trade-off relationship between the fuel consumption rate and the speed error, and that, as a result of the optimization, a favorable parameter set has been successfully obtained as compared to the initial individual. For example, comparing the individuals indicated by dotted circles, the evaluation value of the fuel consumption rate is greatly improved while maintaining or slightly improving the evaluation of the speed error RMS.

  As described above, in the automatic adaptation system 1 of the present embodiment, the influence of the model error or the like in the offline simulation can be eliminated by introducing the actual machine into the learning loop and directly acquiring the evaluation value. It is possible not only to obtain a high parameter, but also to realize a development environment that can reduce the work cost by reducing the burden of parameter adjustment in control development and pursue higher quality performance.

  In particular, in the automatic adaptation system 1 of the present embodiment, by using a multi-objective genetic algorithm that is a multi-objective optimization method, it is not necessary to use a weighting coefficient between a plurality of objective functions, and an optimal solution set is obtained as a Pareto solution set. Can be earned. Thus, by viewing the distribution of the solution set in the evaluation space, the trade-off relationship between the evaluation axes can be easily grasped, and the solution that realizes the most preferable balance of objective functions can be selected.

  Moreover, in the application of this multi-purpose genetic algorithm, the number of individuals is reduced to reduce the number of evaluation value acquisitions, and the number of updates per individual is increased. The time until the optimum value of the control parameter is obtained can be shortened, and an effective control parameter can be acquired within a practical time.

  Also, in the evolution process of multi-objective genetic algorithms, the density of solutions in the evaluation space is calculated, and sparse solutions are preferentially selected as parents (sharing strategy), so even if the number of individuals is small Solution (individuals) can be dispersed, and the trade-off relationship between the evaluation axes can be easily grasped by the finally obtained Pareto solution set, and a variety of control parameters can be obtained. It becomes possible.

  In the determination of the Pareto solution, even if a child individual is determined to be a Pareto solution at the time of the first determination, it is not immediately determined as a Pareto solution. It is possible to obtain a stable Pareto solution by eliminating the variation.

  As a result of re-evaluation of offspring individuals, when variance of evaluation values exceeding the expected variation in evaluation values is detected, the results can be discarded and re-evaluation can be performed to detect and eliminate abnormal values. Become. Moreover, since re-evaluation of the child individual is performed only when it is determined as a Pareto solution at the time of the first determination, the increase in the number of evaluations can be suppressed compared to evaluating all the individuals multiple times, and depending on the actual machine Evaluation time acquisition time can be reduced.

  In addition, since the child individuals are preferentially survived to the next generation in the evolution process, it is possible to follow the drift of the vehicle characteristics and acquire more effective parameters in the actual machine. Furthermore, since the age of the individual is retained and confirmation is performed by re-evaluation with the same control parameters for each determined life, the elimination of the individual whose effectiveness has been lost (corresponding to the Pareto solution certification) corresponding to the characteristic drift of the actual machine Deprivation).

Overall configuration diagram of the automatic calibration system Explanatory drawing showing the speed pattern Block diagram of automatic operation control system Block diagram of learning operation unit Explanatory diagram showing examples of solution groups Illustration of niche size Explanatory drawing showing an example of niche count calculation Flow chart of optimization process Flow chart of offspring evaluation process Explanatory diagram showing an example of the results of child reevaluation Explanatory diagram showing optimization results

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Automatic adaptation system 10 Real machine 20 Automatic driving device 50 Learning calculation part 52 Goodness-of-fit calculation part 53 Parent selection part 54 Generation change part 55 Child individual evaluation part 56 Survival selection part

Claims (6)

An automatic adjustment system for control parameters that adapts control parameters related to control of a moving body to optimum values,
The real machine as the moving body is introduced into the learning loop, the evaluation value is automatically obtained by the automatic operation of the real machine, and the evaluation parameter is independently evaluated by a plurality of evaluation axes, and the control parameter which is the optimum solution is determined by the multi-purpose genetic algorithm With a learning operation unit to be acquired by
The learning calculation unit is
A fitness calculation unit that calculates the fitness of an individual including the control parameter as a gene;
When selecting some individuals as a parent individual from a group of individuals consisting of multiple individuals, if the fitness of the individuals is the same, use an index according to the degree of density with other individuals. A parent selection unit for determining the superiority or inferiority of
A generation alternation part that generates a child individual by genetic manipulation from the selected parent individual,
A child individual evaluation unit that selectively evaluates the generated child individual multiple times;
Have a survival and selection unit for selecting the individual to save the next generation,
The offspring individual evaluation part
An automatic control parameter adaptation system, wherein a re-evaluation of a child individual is performed only when the generated child individual is determined to be a Pareto optimal solution in an initial evaluation . The survival selection unit is
2. The system for automatically adjusting control parameters according to claim 1, wherein when the degree of matching between the generated child individual and the parent individual is equal, the child individual is preferentially stored in the next generation. The survival selection unit is
The individual age is maintained based on the number of generations of the above-mentioned individual, and the evaluation value is confirmed using the control parameter of the individual that has reached the specified life, thereby eliminating the individual whose effectiveness has been reduced. The system for automatically adjusting control parameters according to claim 1. After re-evaluation of said child individual, automatic adaptation system of the control parameters according to claim 1, characterized in that an average value of evaluation values and evaluation values of the child individuals. After re-evaluation of said child individual, when the above value difference or variance is defined between evaluation values, according to claim 1, wherein the discarding part of the evaluation values or the child individual itself as an outlier Automatic adjustment system for control parameters. An automatic adjustment system for control parameters that adapts control parameters related to control of a moving body to optimum values,
The real machine as the moving body is introduced into the learning loop, the evaluation value is automatically obtained by the automatic operation of the real machine, and the evaluation parameter is independently evaluated by a plurality of evaluation axes, and the control parameter which is the optimum solution is determined by the multi-purpose genetic algorithm With a learning operation unit to be acquired by
The learning calculation unit is
A fitness calculation unit that calculates the fitness of an individual including the control parameter as a gene;
When selecting some individuals as a parent individual from a group of individuals consisting of multiple individuals, if the fitness of the individuals is the same, use an index according to the degree of density with other individuals. A parent selection unit for determining the superiority or inferiority of
A generation alternation part that generates a child individual by genetic manipulation from the selected parent individual,
A child individual evaluation unit that selectively evaluates the generated child individual multiple times;
A survival selection unit for selecting individuals to be stored in the next generation,
The parent selection part
The maximum value of the distance between individuals in at least one of the evaluation value space and the control parameter space is defined by either a predesignated value or a ratio to the variance of the evaluation value or the control parameter value. and, automatic adaptation system features and to that control parameter to calculate an index of the degree of the density based on the relation between the distance and the maximum value.

Images