Arithmetic program, arithmetic method, and information processing apparatus

Abstract

To provide an arithmetic program, an arithmetic method, and an information processing apparatus capable of obtaining a robust solution.SOLUTION: An arithmetic program causes a computer to repeatedly execute the processes of: creating an Ising model based on a training data set including a plurality of pieces of training data in which an input variable and an evaluation value are associated with each other; searching for a value of a recommended input variable based on the Ising model; giving a variation to the searched value of the recommended input variable and calculating the evaluation value for the value of the recommended input variable after the variation is given; and adding the value of the recommended input variable and the calculated evaluation value to the training data set as new training data when the value of the recommended input variable does not exist in the training data included in the training data set, and updating a statistical value of the evaluation value of the training data and the evaluation value for the value of the recommended input variable after the variation is given as the evaluation value of the training data when the value of the recommended input variable exists in the training data included in the training data set.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present invention relates to an arithmetic program, an arithmetic method, and an information processing apparatus.

Background Art

[0002] Techniques related to optimization problems have been disclosed (see, for example, Patent Documents 1 to 5).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

Patent Document 5

Summary of the Invention

Problems to be Solved by the Invention

[0004] For example, it is conceivable to perform optimization by sampling using an Ising model in QUBO form. However, it is difficult to obtain a robust solution.

[0005] In one aspect, the present invention aims to provide an arithmetic program, an arithmetic method, and an information processing apparatus capable of obtaining a robust solution.

Means for Solving the Problems

[0006] In one aspect, the arithmetic program causes the computer to perform a process of creating an Ezing model based on a teacher data set including a plurality of teacher data in which input variables and evaluation values are associated, a process of searching for values of recommended input variables based on the Ezing model, a process of giving a variation to the searched values of the recommended input variables and calculating the evaluation value for the values of the recommended input variables after the variation is given, and when the values of the recommended input variables do not exist in the teacher data included in the teacher data set, adding the values of the recommended input variables and the calculated evaluation values as new teacher data to the teacher data set, and when the values of the recommended input variables exist in the teacher data included in the teacher data set, updating the evaluation value of the teacher data with a statistical value between the evaluation value of the teacher data and the evaluation value for the values of the recommended input variables after the variation is given, and repeatedly executing the processes.

Advantages of the Invention

[0007] A robust solution can be obtained.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Embodiments for Carrying Out the Invention

[0009] As a technique for searching for a good solution with a high evaluation value from a large number of combinations, orders, etc., a sampling technique for binary variables is used. Examples of the sampling technique for binary variables include a sampling technique that randomly samples and a sampling technique that uses an Ising model in QUBO format.

[0010] The sampling technique that randomly samples can perform sampling easily, but has the drawback that the sampling efficiency is poor and the number of samplings increases in order to obtain a good solution with high accuracy.

[0011] Therefore, it is conceivable to utilize a sampling technique that uses an Ising model in QUBO format.

[0012] Here, the QUBO format refers to Quadratic Unconstrained Binary Optimization, which is a format that enables binary optimization without quadratic constraints. The QUBO format can be expressed, for example, as follows. Note that x i = 0 or 1 for (i = 1,..., N). W ij is the coupling coefficient between x i and x j b i is the bias coefficient of x i The first term on the right side is a quadratic term and represents an interaction. The second term on the right side is a linear term and represents a bias effect. The third term on the right side is a constant term. In the QUBO format, according to the following formula, a good solution x for minimizing E(x) representing energy is searched for as illustrated in FIG. 1.

Equation

[0013] Optimization problems exist in various industries, including the distribution industry and the manufacturing industry. For these problems, optimization algorithms that seek exact solutions that achieve the highest computational effectiveness have been studied.

[0014] Here, the exact solution refers to the solution that maximizes the evaluation index. For example, in a production line, the evaluation index includes the production completion time, delivery date, production cost, etc. In the packing operation, the evaluation index is the amount of wasted materials, working time, etc. The optimization algorithm is an algorithm that optimizes one or more evaluation indices as the objective function so that the objective function becomes favorable. In a production line, the input variables are the initial input order of products to the production line, as well as conditions such as the working time required for each product. In the packing operation, the input variables are the initial packing order of products, as well as conditions such as the size of each product, the weight of each product, and the size of the box. The obtained solution is the input order of products to the production line, etc. In the packing operation, the obtained solution is the packing order of products, etc.

[0015] With the optimization algorithm, an exact solution that maximizes the evaluation index can be obtained. However, when attempting to actually apply the exact solutions obtained by these optimization algorithms, various fluctuations can occur in reality. Therefore, it is difficult to perform operations exactly as per the exact solution, or the physical properties cannot be obtained as per the exact solution, and as a result, it is considered that there will be a slight deviation from the exact solution.

[0016] Here, the exact solutions obtained by conventional methods are often solutions that achieve a locally high effect. Therefore, if the work content and physical properties deviate from the exact solution (= a tolerance occurs), it is considered that the effect will sharply decrease and the expected effect cannot be obtained. Also, even if the work content is as per the exact solution, since the variable factors at the site are not taken into account, it is considered that the expected effect cannot be obtained in reality.

[0017] For example, in the packing operation, events such as the required time of the operation becoming longer due to differences in the proficiency of the workers, or the inability to perform the packing operation exactly as per the exact solution due to size errors of the products handled in the operation, etc. can be considered. These are things that can occur in the actual field. If these events occur, the work content will deviate from the exact solution.

[0018] Figure 2 is a diagram illustrating the optimal solution. In Figure 2, the x-axis shows the value obtained by converting the 0 and 1 bit arrays represented in Figure 1 into numerical values from 0 to 1. The y-axis shows the objective function, corresponding to E(x) in Figure 1. In Figure 2, it is shown that the smaller the value of y, the better the result obtained.

[0019] For example, the optimal solution around x = 0.1 is the exact solution because the value of y is the smallest. In this exact solution, although the value of y is the smallest, since the solution exists locally, even if the value of x deviates slightly from the exact solution, the value of y increases significantly. This is because the width in the horizontal axis direction of the valley near the exact solution in the graph of Figure 2 is small. On the other hand, for the optimal solution around x = 0.5, the value of y is not as small as the exact solution, but compared to the exact solution, the solution does not exist locally. That is, in the graph of Figure 2, the width in the horizontal axis direction of the valley near the optimal solution around x = 0.5 is larger than the width in the horizontal axis direction of the valley near the optimal solution (exact solution) around x = 0.1. Therefore, even if the work content deviates slightly from the optimal solution, the value of y does not increase significantly. In the actual field, even when the work content deviates from the optimal solution, a highly robust optimization algorithm that can obtain an effect close to the optimal solution is required.

[0020] However, in optimization problems, there are problems that are complex, difficult to formulate as objective functions, and have high computational costs. Therefore, based on the input and output data of the optimization problem, it is conceivable to use a black-box optimization method that repeatedly performs a process of setting an input as a new search point, obtaining its output, and then newly searching based on the result to search for the optimal input. As one such optimization method, it is conceivable to use an FM (machine learning method) model with low computational cost.

[0021] For example, FMQA (Factorization Machine with Quantum Annealing) etc. can be mentioned. FMQA is a method that combines QA (Quantum Annealing) and FM. In FMQA, an FM model in QUBO format is created from training data, a good solution is obtained by QA, the evaluation value of the good solution is analyzed by a property solver, the result is added to the training data, and sampling is performed interactively. As other sampling techniques using QUBO format models, the FMDA technique etc. can be mentioned. FMDA is obtained by replacing the QA part of FMQA with DA (Digital Annealer).

[0022] Here, the outline of FMQA will be explained. Fig. 3 is a flowchart showing the repetition unit in FMQA. First, a training dataset having a plurality of data of input variables and evaluation values is prepared. At the first execution, the training dataset is prepared randomly. The input variable is the value of the horizontal axis x explained in Fig. 2. For example, a plurality of values are randomly generated in the range of x = 0 to 1.27. The evaluation value is the evaluation value calculated by the property solver for each input variable.

[0023] Next, for each piece of training data included in the training dataset, the input variable is converted into a bit array. For example, the decimal value obtained by multiplying the input variable from 0 to 1.27 by 100 is converted into a 7-bit binary number. As a result, the input variable can be converted into a vector format of x = (x1, x2, …, x M ) where each value is a binary value of 0 or 1.

[0024] Next, an FM model is created using the teacher dataset and converted into QUBO to create an Ising model in QUBO format. Since FM is already in QUBO format, creating an FM is equivalent to creating an Ising model in QUBO format. Any other machine learning model can be used as long as it can create an Ising model in QUBO format.

[0025] Next, using QA, the created Ising model in QUBO format is optimized to generate a good solution (recommended bit array) with the best evaluation value.

[0026] Next, the recommended input variables are calculated by converting the recommended bit array back to numerical values in the range of x = 0 to 1.27.

[0027] Next, the recommended input variables are analyzed by the characteristic solver to calculate the evaluation value.

[0028] Next, the evaluation results (the set of recommended input variables and evaluation values) are newly added to the teacher dataset as teacher data. By repeating these series of processes, the teacher dataset is updated and the optimal solution can be obtained.

[0029] In Fig. 3, the QA part can be replaced with DA or the like, and any other method can be used as long as it is an Ising machine that can solve QUBO. In addition to using a dedicated Ising machine for solving QUBO, the Ising model can also be solved using software.

[0030] Such sampling techniques are methods for solving the exact solution with the minimum evaluation value. Therefore, a robust solution considering fluctuations is not necessarily obtained.

[0031] Therefore, as a robust optimization method, it is conceivable to use the MORDO (multi-objective robust design optimization) method. This method is a technique that repeatedly samples and varies the design variables to be searched during the optimization calculation, and statistically processes and evaluates the obtained indicators. The average value and standard deviation value of the objective function are used as evaluation values, and finally, a solution with a high evaluation value is calculated as the optimal solution. In the MORDO method, since a sufficient number of sampling points are repeatedly evaluated in the vicinity of the input variables, a highly robust solution can be obtained. However, when using a characteristic solver with a large computational cost, the computational cost increases and it becomes unrealistic.

[0032] Therefore, in the following examples, an example that can obtain a robust solution while suppressing the computational cost will be described.

Example

[0033] First, the principle of this example will be described. FIG. 4 is a diagram for explaining the principle of this example. As illustrated in FIG. 4, FMQA as described in FIG. 3 is used. However, the recommended input variables obtained by QA are varied, and the evaluation values are calculated using the characteristic solver. Next, a set of the recommended input variables and the evaluation values calculated by varying the recommended input variables is added to the teacher dataset as new teacher data. As a result, input variables having a plurality of different evaluation values can be obtained. Therefore, statistical values of the evaluation values (for example, average value, weighted average value, etc.) are calculated for each input variable. For input variables having a plurality of different evaluation values, this statistical value is used as the evaluation value.

[0034] For example, the teacher dataset is illustrated in FIG. 5(a). As illustrated in FIG. 5(a), assume that the teacher dataset includes input variables x1, x2, x3,.... Also, assume that the evaluated value for the input variable x1 is f(x1), the evaluated value for the input variable x2 is f(x2), and the evaluated value for the input variable x3 is f(x3). In this case, assume that the recommended input variable is x1 and a variation α is given. In this case, since f(x1 + α) is calculated as the evaluated value, there are two evaluated values for the input variable x1, namely f(x1) and f(x1 + α).

[0035] Therefore, as illustrated in FIG. 5(b), ((f(x1) + f(x1 + α)) / 2), which is the average value of these evaluated values, is updated as the evaluated value for the input variable x1. When there is no recommended input variable in the teacher dataset, the recommended input variable and the evaluated value calculated by applying a variation to the input variable are added to the teacher dataset as new teacher data.

[0036] When the process of FIG. 4 is repeated, the teacher dataset is updated by applying a variation to the recommended input variable recommended as the optimal solution with the best evaluated value. As a result, it becomes possible to construct a model that averages the variation in the evaluated value due to the applied variation.

[0037] For example, when the exact solution described in FIG. 2 becomes the recommended input variable, the evaluated value deteriorates due to the applied variation. As a result, it becomes difficult for the input variable near the exact solution to be recommended as the optimal solution. On the other hand, when the robust solution described in FIG. 2 becomes the recommended input variable, the evaluated value remains good even when a variation is applied. Therefore, it becomes easy for the input variable near the robust solution to be recommended as the optimal solution. As a result, as illustrated in FIG. 5(c), it becomes possible to obtain a solution with high robustness and an average small variation in the evaluated value due to the variation.

[0038] Hereinafter, the apparatus configuration for realizing the above principle will be described. FIG. 6(a) is a block diagram illustrating the overall configuration of the information processing apparatus 100. As illustrated in FIG. 6(a), the information processing apparatus 100 includes a storage unit 10, an initial point generation unit 20, an evaluation unit 30, an FMQA execution unit 40, a teacher data update unit 50, an output unit 60, and the like.

[0039] FIG. 6(b) is a block diagram illustrating the hardware configuration of the information processing apparatus 100. As illustrated in FIG. 6(b), the information processing apparatus 100 includes a CPU 101, a RAM 102, a storage device 103, an input device 104, a display device 105, and the like.

[0040] The CPU (Central Processing Unit) 101 is a central arithmetic processing unit. The CPU 101 includes one or more cores. The RAM (Random Access Memory) 102 is a volatile memory that temporarily stores programs executed by the CPU 101, data processed by the CPU 101, and the like. The storage device 103 is a non-volatile storage device. As the storage device 103, for example, a solid-state drive (SSD) such as a ROM (Read Only Memory), a flash memory, a hard disk driven by a hard disk drive, or the like can be used. The storage device 103 stores an arithmetic program. The input device 104 is an input device such as a keyboard or a mouse. The display device 105 is a display device such as an LCD (Liquid Crystal Display). By the CPU 101 executing the arithmetic program, the storage unit 10, the initial point generation unit 20, the evaluation unit 30, the FMQA execution unit 40, the teacher data update unit 50, the output unit 60, and the like are realized. Note that, as the storage unit 10, the initial point generation unit 20, the evaluation unit 30, the FMQA execution unit 40, the teacher data update unit 50, the output unit 60, and the like, hardware such as a dedicated circuit may be used.

[0041] FIG. 7 is a flowchart showing an example of the operation of the information processing apparatus 100. As illustrated in FIG. 7, the initial point generation unit 20 sets the original data of the initial teacher data and stores it in the storage unit 10 (step S1).

[0042] Next, the evaluation unit 30 calculates the average value of the evaluation values for each input variable (step S2). When step S2 is executed for the first time, since a plurality of evaluation values have not yet been obtained for the input variable, each evaluation value is treated as the average value.

[0043] Next, the teacher data update unit 50 associates the bit array with the average value of the evaluation values for the input variable for which the average value of the evaluation values has been calculated, and sets it as teacher data (step S3). When step S3 is executed for the first time, the original data in step S1 becomes the teacher data.

[0044] Next, the FMQA execution unit 40 creates an FM model using the teacher data obtained in step S3 (step S4). Since the FM is in the QUBO format, creating the FM is equivalent to creating an Ising model in the QUBO format. Any other machine learning model that can create an Ising model in the QUBO format can be used.

[0045] Next, the FMQA execution unit 40 generates the recommended points (recommended bit arrays) for the set number of recommendations using the FM created in step S4 (step S5). The set number of recommendations is set in advance by the user. The recommended point is one optimal solution (recommended point) that maximizes the evaluation value. Or, the recommended point is one or more optimal solutions (recommended points) whose evaluation values are equal to or greater than the threshold value. Or, the recommended point is a plurality of optimal solutions (recommended points) up to the predetermined rank in terms of the evaluation value.

[0046] Next, the FMQA execution unit 40 converts the recommended bit array generated in step S5 into a recommended input variable (step S6).

[0047] Next, the FMQA execution unit 40 determines whether the convergence condition is satisfied (step S7). For example, the FMQA execution unit 40 determines whether the number of iterations has reached the upper limit. The number of executions of step S7 may be regarded as the number of iterations. The upper limit number of iterations is set in advance by the user. If it is determined "Yes" in step S7, the execution of the flowchart ends.

[0048] If it is determined "No" in step S7, the FMQA execution unit 40 adds a variation to the recommended input variable (step S8).

[0049] Next, the evaluation unit 30 calculates an evaluation value for the recommended input variable after the variation is added, using a solver (step S9).

[0050] Next, the teacher data update unit 50 adds the recommended input variable before the variation is added and the evaluation value calculated in step S9 to the teacher data set as teacher data (step S10). Then, the process is executed again from step S2.

[0051] The output unit 60 outputs the result of the process in FIG. 7. The output result is displayed, for example, by the display device 105. For example, the output unit 60 may output the content of the teacher data set, or may output the teacher data with a high evaluation value in the teacher data set as the optimal solution.

[0052] According to this embodiment, a variation is given to the searched recommended input variable, and an evaluation value for the recommended input variable after the variation is given is calculated. When there is no recommended input variable in the teacher data included in the teacher data set, the recommended input variable and the calculated evaluation value are added to the teacher data set as new teacher data. When there is a recommended input variable in the teacher data included in the teacher data set, the statistical value of the evaluation value of the teacher data and the evaluation value of the recommended input variable after the variation is given is updated as the evaluation value of the teacher data. By updating the teacher data set in this way, a robust solution can be obtained.

[0053] In the above example, the variable value = α given to the recommended input variable is one, but it may be two or more. For example, for the recommended input variable x1, variable values = α1 and variable value = α2 may be given. In this case, since the evaluation values f(x1), f(x1 + α1), and f(x1 + α2) can be obtained for the recommended input variable x1, statistical values such as the average value of these evaluation values may be calculated. However, since the calculation cost of the characteristic solver is high, it is preferable that the variable value given to the recommended input variable is one.

[0054] Also, if the variable value is too large, there is a risk that the evaluation value will deteriorate even for the robust solution. Therefore, an upper limit may be set for the variable value. For example, the upper limit of the variable value can be set to a fixed value. For example, since the upper limit of the variable value can be changed by the user, it can be set to an upper limit value according to the optimization problem. On the other hand, if the variable value is too small, there is a risk that the probability of searching for the exact solution as the optimal solution will increase. Therefore, a lower limit may be set for the variable value. For example, the lower limit of the variable value can be set to a fixed value. For example, since the lower limit of the variable value can be changed by the user, it can be set to a lower limit value according to the optimization problem.

[0055] (Simulation results) Hereinafter, a virtual problem will be set, and the simulation results of performing the arithmetic processing according to the above embodiment will be described. Specifically, the maximization problem of the formula f b (x) which is the standard problem of the following formula (2) is treated as a minimization problem, and the minimization problem of the following formula (3) is solved. Note that the following formula (3) is represented in FIG. 8. The exact solution is around x = 0.1, and the robust solution is around x = 0.5.

Number

Number

[0056] For each value of x from 0 to 1.27, the decimal value multiplied by 100 was converted into a 7-bit binary number. The FM model of the following formula (4) was used. In the FM model, k = 3, and the initial teacher data was 10 random data. The number of iterations (number of repetitions) was set to 300. As the optimization engine, a single-objective GA was used. In the GA, the number of evolutionary generations was set to 50 generations, and the number of individuals was set to 10 individuals. The initial values were the same 10 random data as the initial teacher data of the FM model. When varying the recommended input variables, white noise in the range of [-0.05, 0.05] was set as random numbers at the time of input to the characteristic solver.

Number

[0057] In each of the upper, middle, and lower graphs of Fig. 9, the horizontal axis represents the input variable x. In the upper and lower graphs of Fig. 9, the vertical axis represents fb´(x). In the middle graph of Fig. 9, the vertical axis represents the number of iterations. In the upper graph of Fig. 9, "Objective function" represents the minimization problem shown in Fig. 8. "Optimal Solution" represents the optimal solution. "300th FM model" represents the FM model obtained by 300 repetitions. In the lower graph of Fig. 9, "300th training data" represents the 300th teacher data, and "Average of 300th training data" represents the average value of the 300th teacher data.

[0058] As shown in the middle graph of Fig. 9, when the number of iterations is small, around x = 0.1 is recommended as the recommended input variable, but as the number of iterations increases, around x = 0.5 is recommended as the recommended input variable. Therefore, it can be seen that it is possible to obtain a robust solution.

[0059] In the above example, the FMQA execution unit 40 creates an aging model based on a teacher dataset including a plurality of pieces of teacher data in which input variables and evaluation values are associated, searches for recommended input variables for the aging model, and when the recommended input variables do not exist in the teacher data included in the teacher dataset, adds the recommended input variables and the evaluation values calculated by varying the recommended input variables as new teacher data to the teacher dataset, and when the recommended input variables exist in the teacher data included in the teacher dataset, updates the evaluation value of the teacher data with a statistical value of the evaluation value of the teacher data and the evaluation value calculated for the recommended input variables after variation. This corresponds to an execution unit that repeatedly executes the above processes.

[0060] As described above, the embodiments of the present invention have been described in detail. However, the present invention is not limited to such specific embodiments, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims. (Appendix 1) Figure 7 On a computer, a process of creating an aging model based on a teacher dataset including a plurality of pieces of teacher data in which input variables and evaluation values are associated, a process of searching for values of recommended input variables based on the aging model, a process of varying the values of the searched recommended input variables and calculating the evaluation values for the values of the recommended input variables after variation, when the values of the recommended input variables do not exist in the teacher data included in the teacher dataset, adding the values of the recommended input variables and the calculated evaluation values as new teacher data to the teacher dataset, and when the values of the recommended input variables exist in the teacher data included in the teacher dataset, updating the evaluation value of the teacher data with a statistical value of the evaluation value of the teacher data and the evaluation value for the values of the recommended input variables after variation, and repeatedly executing the above processes. This is a characteristic of an arithmetic program. (Appendix 2) The calculation program according to Appendix 1, wherein the statistical value is an average value. (Appendix 3) The calculation program according to Appendix 1, wherein when varying the value of the recommended input variable, there is one variation value. (Appendix 4) The calculation program according to Appendix 3, wherein the variation value is a random value using a random number. (Appendix 5) The calculation program according to Appendix 1, wherein when varying the value of the recommended input variable, an upper limit is set for the variation value. (Appendix 6) The calculation program according to Appendix 1, wherein when creating the aging model, the input variable of the teacher data is converted into binary vector data. (Appendix 7) A computer a process of creating an aging model based on a teacher data set including a plurality of pieces of teacher data in which an input variable and an evaluation value are associated; a process of searching for a value of a recommended input variable based on the aging model; a process of varying the searched value of the recommended input variable and calculating an evaluation value for the value of the recommended input variable after the variation; When the value of the recommended input variable does not exist in the teacher data included in the teacher data set, the value of the recommended input variable and the calculated evaluation value are added to the teacher data set as new teacher data, and when the value of the recommended input variable exists in the teacher data included in the teacher data set, a statistical value between the evaluation value of the teacher data and the evaluation value for the value of the recommended input variable after the variation is updated as the evaluation value of the teacher data, and the processes are repeatedly executed. A calculation method characterized by this. (Appendix 8) The calculation method according to Appendix 7, wherein the statistical value is an average value. (Appendix 9) The calculation method according to Appendix 7, wherein when varying the value of the recommended input variable, there is one variation value. (Appendix 10) The calculation method according to Appendix 9, wherein the variable value is a random value using a random number. (Appendix 11) The calculation method according to Appendix 7, wherein when varying the value of the recommended input variable, an upper limit is set for the variable value. (Appendix 12) The calculation method according to Appendix 7, wherein when creating the aging model, the input variable of the teacher data is converted into binary vector data. (Appendix 13) A process of creating an aging model based on a teacher data set including a plurality of teacher data in which an input variable and an evaluation value are associated, a process of searching for a value of a recommended input variable based on the aging model, a process of varying the searched value of the recommended input variable and calculating an evaluation value for the value of the recommended input variable after the variation, and when the value of the recommended input variable does not exist in the teacher data included in the teacher data set, adding the value of the recommended input variable and the calculated evaluation value as new teacher data to the teacher data set, and when the value of the recommended input variable exists in the teacher data included in the teacher data set, updating a statistical value between the evaluation value of the teacher data and the evaluation value for the value of the recommended input variable after the variation as the evaluation value of the teacher data, and an execution unit that repeatedly executes the above processes. An information processing apparatus characterized by comprising: (Appendix 14) The information processing apparatus according to Appendix 13, wherein the statistical value is an average value. (Appendix 15) The information processing apparatus according to Appendix 13, wherein the execution unit makes the variable value when varying the value of the recommended input variable to be one. (Appendix 16) The information processing apparatus according to Appendix 15, wherein the execution unit uses a random value using a random number as the variable value. (Appendix 17) The information processing apparatus according to Supplementary Note 13, wherein the execution unit sets an upper limit for the variation value when varying the value of the recommended input variable. (Supplementary Note 18) The information processing apparatus according to Supplementary Note 13, wherein the execution unit converts the input variable of the teacher data into binary vector data when creating the aging model.

Explanation of Signs

[0061] 10 Storage unit 20 Initial point generation unit 30 Evaluation unit 40 FMQA execution unit 50 Teacher data update unit 60 Output unit 100 Information processing apparatus 101 CPU 102 RAM 103 Storage device 104 Input device 105 Display device

Claims

1. A computer, a process of creating an EDM model based on a teacher data set including a plurality of teacher data in which an input variable and an evaluation value are associated; a process of searching for a value of a recommended input variable based on the EDM model; a process of applying a variation to the searched value of the recommended input variable and calculating the evaluation value for the value of the recommended input variable after the variation is applied; when the value of the recommended input variable does not exist in the teacher data included in the teacher data set, adding the value of the recommended input variable and the calculated evaluation value as new teacher data to the teacher data set, and when the value of the recommended input variable exists in the teacher data included in the teacher data set, updating the evaluation value of the teacher data with a statistical value between the evaluation value of the teacher data and the evaluation value for the value of the recommended input variable after the variation is applied, and repeatedly executing the processes, an arithmetic program characterized by that.

2. The arithmetic program according to claim 1, wherein the statistical value is an average value.

3. The arithmetic program according to claim 1, wherein there is one variation value when applying a variation to the value of the recommended input variable.

4. The arithmetic program according to claim 3, wherein the variation value is a random value using a random number.

5. The arithmetic program according to claim 1, wherein an upper limit is set for the variation value when applying a variation to the value of the recommended input variable.

6. The arithmetic program according to claim 1, wherein when creating the EDM model, the input variable of the teacher data is converted into binary vector data.

7. A computer, a process of creating an EDM model based on a teacher data set including a plurality of teacher data in which an input variable and an evaluation value are associated; a process of searching for a value of a recommended input variable based on the EDM model; a process of applying a variation to the searched value of the recommended input variable and calculating the evaluation value for the value of the recommended input variable after the variation is applied; When the value of the recommended input variable does not exist in the teacher data included in the teacher dataset, the value of the recommended input variable and the calculated evaluation value are added to the teacher dataset as new teacher data. When the value of the recommended input variable exists in the teacher data included in the teacher dataset, a process of updating the evaluation value of the teacher data with a statistical value between the evaluation value of the teacher data and the evaluation value for the value of the recommended input variable after variation is repeatedly executed. A calculation method characterized by this.

8. A process of creating an Ezing model based on a teacher dataset including a plurality of teacher data in which input variables and evaluation values are associated, a process of searching for a value of a recommended input variable based on the Ezing model, a process of applying a variation to the searched value of the recommended input variable and calculating an evaluation value for the value of the recommended input variable after the variation, and when the value of the recommended input variable does not exist in the teacher data included in the teacher dataset, the value of the recommended input variable and the calculated evaluation value are added to the teacher dataset as new teacher data. When the value of the recommended input variable exists in the teacher data included in the teacher dataset, a process of updating the evaluation value of the teacher data with a statistical value between the evaluation value of the teacher data and the evaluation value for the value of the recommended input variable after variation is repeatedly executed. An information processing apparatus characterized by including an execution unit for this.

Images