Information processing device, information processing method, and recording medium

Abstract

An information processing device 100 according to the present disclosure comprises: a learning unit 121 that, when learning and generating an approximate function for approximating distribution of a data set including an explanatory variable and an objective variable, learns and generates a plurality of approximate functions from the data set by using a loss function based on the degree of convexity set for the approximate function; and a calculation unit 122 that obtains a solution for the objective variable on the basis of the plurality of approximate functions.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to an information processing device, an information processing method, and a recording medium.

[0002] Functions generated by machine learning are used to perform optimization so as to satisfy set constraints. For example, this can be used to optimize the motion planning of a robot, or to optimize factory production planning as described in Patent Literature 1.

[0003] JP 2019-124990 A

[0004] However, the optimization method described above requires sampling variables such as plans until the value of the learned function satisfies the constraints, which increases the computational cost and time, resulting in problems such as delays in the control of robots or factories.

[0005] Therefore, one object of the present disclosure is to provide an information processing device that can reduce the delay in the control described above, which is the problem described above.

[0006] An information processing device according to one aspect of the present disclosure includes: a learning unit that, when learning and generating approximation functions that approximate a distribution of a dataset including explanatory variables and a dependent variable, learns and generates a plurality of the approximation functions from the dataset using a loss function based on a degree of convexity set for the approximation functions; and a calculation unit that finds a solution for the dependent variable based on the plurality of approximation functions.

[0007] Furthermore, an information processing method according to one aspect of the present disclosure is configured to: when learning and generating an approximation function that approximates a distribution of a dataset including explanatory variables and a dependent variable, learn and generate a plurality of the approximation functions from the dataset using a loss function based on a degree of convexity set for the approximation function; and obtain a solution for the dependent variable based on the plurality of the approximation functions.

[0008] Furthermore, a program according to one aspect of the present disclosure is configured to execute, on a computer, the following processing: when learning and generating an approximation function that approximates the distribution of a dataset including explanatory variables and a dependent variable, learning and generating a plurality of the approximation functions from the dataset using a loss function based on a degree of convexity set for the approximation functions; and obtaining a solution for the dependent variable based on the plurality of the approximation functions.

[0009] With the above-described configuration, the present disclosure can reduce delays in control and the like.

[0010] FIG. 1 is a block diagram showing the configuration of an information processing device according to a first embodiment of the present disclosure. FIG. 2 is a diagram for explaining processing by the information processing device disclosed in FIG. 1. FIG. 3 is a diagram showing the state of processing by the information processing device disclosed in FIG. 1. FIG. 4 is a diagram showing the state of processing by the information processing device disclosed in FIG. 1. FIG. 5 is a flowchart showing the operation of the information processing device disclosed in FIG. 1. FIG. 6 is a block diagram showing the configuration of an information processing device according to a second embodiment of the present disclosure. FIG. 7 is a block diagram showing the hardware configuration of an information processing device according to a third embodiment of the present disclosure. FIG. 8 is a block diagram showing the configuration of an information processing device according to the third embodiment of the present disclosure.

[0011] First Embodiment A first embodiment of the present disclosure will be described with reference to Fig. 1 to Fig. 6. Fig. 1 is a diagram for explaining the configuration of an information processing device, and Fig. 2 to Fig. 6 are diagrams for explaining the processing operation of the information processing device.

[0012] [Configuration] The information processing device 10 of this embodiment is suitable as an optimization device that uses a function generated by machine learning or the like to optimize a system to satisfy set constraints. Examples of optimization processes performed by the information processing device 10 include optimization of a robot's motion plan and optimization of a factory's production plan. Specifically, optimization of a robot's motion plan involves learning a function that uses the robot's control and the object's position and orientation as inputs, outputs the feasibility of the motion, and then constraining the output to be equal to or greater than a certain value, thereby determining the robot's motion plan. However, the optimization process performed by the information processing device 10 may be used to optimize any target. Hereinafter, for convenience, the term "optimization" will be used to describe the processing performed by the information processing device 10, but it is not necessary to find a point where the function value is minimum. For example, the information processing device 10 may perform a search process to decrease the function value and find a point where the search result satisfies a criterion for determining that the search result is a solution. Examples of criteria for determining that the search result is a solution include a criterion that the number of iterations reaches a predetermined number of iterations, a criterion that the function value is equal to or less than a predetermined value, or a criterion that a point where the function value decreases is not found within a predetermined number of iterations. Furthermore, the information processing device 10 may have a function of controlling a control target such as a robot, a manufacturing factory, a railway, a chemical plant, an agricultural plant, a logistics warehouse, etc. in accordance with the obtained solution. In this case, it can also be said that the information processing device 10 functions as a control device.

[0013] The information processing device 10 is composed of one or more information processing devices each including a calculation device and a storage device. As shown in FIG. 1, the information processing device 10 includes a learning unit 11 and an optimization unit 12. The functions of the learning unit 11 and the optimization unit 12 can be realized by the calculation device executing a program for realizing each function stored in the storage device. The information processing device 10 also includes a learning data storage unit 16. The learning data storage unit 16 is composed of a storage device. Each component will be described in detail below.

[0014] The training data storage unit 16 stores training data, which is a data set consisting of explanatory variables and objective variables acquired from the optimization target. The training data may be a data set acquired by measuring the target or its simulator, or may be a data set used when generating the training model.

[0015] The learning unit 11 learns and generates an approximation function that approximates the distribution of the training data. At this time, the learning unit 11 learns and generates multiple approximation functions from the training data using a loss function based on the degree of convexity of the approximation function. In other words, the learning unit 11 learns and generates multiple approximation functions so that the approximation functions gradually approximate convex functions. The process of generating the approximation functions by the learning unit 11 will be described in further detail below.

[0016] First, a convex function will be briefly explained with reference to Fig. 2. A convex function is a function with few local solutions, and the following formula (1) shown in Fig. 2 is a function that is expressed in the interval [x １ , x ２ ], if it holds for any t (where 0≦t≦1), then the function f is in the interval [x １ , x ２ ] is downward convex.

[0017] Here, in this embodiment, the processing will be described using an example in which the function is downwardly convex, but an upwardly convex function may also be used. In this case, the information processing device 10 executes a process in which the sign of the process described below is inverted, thereby finding the optimal value. Here, a discrete convexity (temperature) parameter α is set as follows: α=0, 0.1, 0.2, ..., 0.9, 1.0 (0≦α≦1). However, the above α is merely an example, and the information processing device 10 may execute the processing described below for a plurality of α that satisfy 0≦α≦1. Using this convexity parameter α, the loss function L α and set the approximate function y = f α (x) are generated by learning.

[0018] In this case, the smaller the value of α, the smaller the approximate function f generated by learning, as will be described later. α(x) is highly approximate to a convex function, that is, the approximate function f α (x) has a high degree of convexity. α The closer (x) is to a convex function, the closer the approximation function f α Therefore, in this embodiment, the approximation function f α Approximating (x) to a convex function means approximating (regressing) with a function that is close to a certain degree to the function f(x) that represents the distribution of the original training data and has few local solutions. For example, when α is close to 0, the approximate function f α Since the degree of convexity of (x) is high and the degree of approximation to a convex function is high, the number of local solutions becomes smaller. On the other hand, as α approaches 1, the approximation function f α The degree of convexity of (x) becomes low, and the approximation function f α (x) has a high degree of approximation to the distribution of the training data, and has a shape that is fitted by such distribution, which may result in more local solutions.

[0019] And the loss function L α is expressed by the following formula 2.

[0020] Here, the loss function L shown in the first term on the right side of Equation 2 １ is expressed by the approximate function f α (x) represents the loss function for regression. Also, the loss function L shown in the second term on the right side of Equation 2 ２ is expressed by the approximate function f α represents the loss function for regression of the mean value of (x). Note that the denominator |α| in Equation 4 contains the number of divisions of α. In other words, in the above-mentioned case of "α = 0, 0.1, 0.2, ..., 0.9, 1.0", the number of α is "11".

[0021] In addition, the loss function L shown in the third term on the right side of Equation 2 α ｃｏｎｖｅｘ is expressed by the following formula 5. Note that (1-α) in formula 5 ２The term is a square, but it may be any value such as a first power or a third power, and may be a function that monotonically decreases as α changes from 0 to 1. α is an appropriate (x １ , x ２ , t). For example, if the search space is 0≦x≦10 and the search space is divided into 10, then (x １ , x ２ ) can be, for example, (0,1), (1,2),..., (8,9), (9,10). １ , x ２ Similarly, for t, the value when appropriately divided in the range of 0≦x≦1 (for example, t=0, 0.1, 0.2, ..., 0.9, 1.0) is C α Included in. In this case, I in Equation 5 ｃ is expressed by the following equation 6.

[0022] The right side of the above equation (6) is obtained by moving the right side of the equation (1), which is an inequality indicating that the function is a convex function, to the left side, and represents the convexity of the function. ｃ ≦0, and max(I ｃ , 0) becomes "0", and the equation 5 itself becomes "0". In this case, since the function is convex, the loss function L α ｃｏｎｖｅｘ On the other hand, the constraint by I ｃ If >0, then max(I ｃ , 0) is "I ｃ ", and the smaller the value of Ic, the smaller the value of L α ｃｏｎｖｅｘ changes depending on the value of α. Therefore, the loss function L α ｃｏｎｖｅｘ By setting each of these, it is possible to set a loss function according to the target convexity degree. In particular, since (1-α) is used in Equation 5, the closer the value of α is to 0, that is, the greater the target convexity degree, the greater the loss function L α ｃｏｎｖｅｘbecomes a large value, and the closer the value of α is to 1, that is, the smaller the target degree of convexity, the smaller L α ｃｏｎｖｅｘ becomes a small value.

[0023] Then, as described above, the learning unit 11 calculates the loss function L shown in Equation 2, which is set for each value of the convexity parameter α. α Using this, the approximate function f of the distribution of the training data α (x) is learned and generated. At this time, the loss function L α Among them, the loss function L α ｃｏｎｖｅｘ Regarding the term, the smaller the value of α, that is, the larger the target degree of convexity, the greater the α ｃｏｎｖｅｘ Since the value of becomes large, minimizing this value is emphasized in learning, and as a result, the approximation function f α (x) can be generated. Conversely, the loss function L α ｃｏｎｖｅｘ Regarding the term, the larger the value of α, that is, the smaller the target degree of convexity, the smaller L α ｃｏｎｖｅｘ Since the value of is small, the effect of the convex constraint is low, and the approximation function f α For example, when α=1.0, the loss function L α ｃｏｎｖｅｘ Since the term is "0", it is the same as learning to find a normal approximate function.

[0024] 3 to 5 show the approximate function f generated by changing the value of the convexity parameter α as described above. α In this example, the convexity parameter α is changed to 0, 1 / 9, 2 / 9, ..., 8 / 9, 1, and the approximate function f generated by learning in each case is shown. α (x) is shown by a solid line. The distribution of the training data is shown by dots.

[0025] As shown in FIG. 3, when the value of the convexity parameter α is smaller, the approximation function f αAs shown in FIGS. 4 and 5, the larger the value of the convexity parameter α, the more closely the approximate function f(x) approaches the distribution of the training data and can match it, and the more closely the convexity of the approximate function f(x) is generated. α (x) is generated.

[0026] The optimization unit 12 (calculation unit) calculates the plurality of approximate functions f α (x) to search for an optimal solution and perform optimization. In particular, in this embodiment, the optimization unit 12 sequentially searches for the corresponding approximation function f α (x) is used to search for a solution and perform optimization. That is, the approximate function f shown in FIGS. α In the example of (x), the approximate functions f α (x). At this time, the optimization unit 12 searches for a solution to the approximate function f α In (x), the previous order approximation function f α The solution obtained from (x) is applied to the later order approximation function f α The solution search is performed using the initial solution of (x). In this embodiment, the solution search is performed by searching for a local solution using the gradient descent method, for example.

[0027] Specifically, an example of searching for a solution in Fig. 3 to Fig. 5 will be described. In Fig. 3 to Fig. 5, solutions are indicated by stars. First, the approximate function f α (x) is used to search for a solution. At this time, the approximate function f α Since (x) has a high degree of convexity, it is easy to search for a solution using the gradient descent method. Next, the approximate function f α (x) is searched for a solution. At this time, the previous α=0 approximation function f α The solution at (x) is expressed as an approximation function f α (x) is set as the initial solution, and a solution search is performed from this initial solution using the gradient descent method. In this way, the search can be started from the vicinity of the initial solution, and since the degree of convexity is high, the solution search becomes easy. Then, furthermore, the approximate function fα (x) is searched for a solution. At this time, the previous α=1 / 9 approximation function f α The solution at (x) is expressed as an approximation function f α (x) is set as the initial solution, and a solution search is performed. In this way, by repeating the solution search sequentially up to α=1, the finally found solution can be determined as the optimal solution.

[0028] [Operation] Next, the operation of the information processing device 10 described above will be described mainly with reference to the flowchart of FIG.

[0029] First, the information processing device 10 learns and generates an approximation function that approximates the distribution of the learning data (step S1). α Using a loss function based on the degree of convexity of (x), multiple approximate functions f are obtained from the training data. α For example, for each degree of convexity expressed using a discrete convexity parameter α, a plurality of approximate functions f as shown in FIGS. α Here, the lower the value of the convexity parameter α, the higher the degree of convexity.

[0030] The information processing device 10 generates a plurality of approximate functions f α (x) is used to perform optimization. At this time, the information processing device 10 first calculates the unprocessed approximate function f α The approximate function f with the highest degree of convexity among (x) α (x) is selected (step S2), and the approximation function f α (x) (step S3). After that, the unprocessed approximate function f α If any (x) remains (Yes in step S4), the next most convex approximation function f α (x) is selected (step S2), and the approximation function f α A solution is searched for from (x) (step S3). At this time, the approximate function f α The search for a solution in (x) is performed starting from a previously searched solution, which is used as an initial solution.

[0031] And the unprocessed approximation function f α Each of the above-mentioned approximate functions f α The search for a solution in (x) is repeated (No in step S4), and the final approximate function f α The solution to (x) is calculated as the optimal solution (step S5). In other words, a solution for the dependent variable is obtained using an approximation function for a data set including the explanatory variables and the dependent variable.

[0032] As described above, in this embodiment, multiple approximation functions with different degrees of convexity are learned, and solutions are obtained using these approximation functions for optimization. This makes it easier to search for a solution for an approximation function with a high degree of convexity. Furthermore, by using a solution for an approximation function with a high degree of convexity as the initial solution for an approximation function with a low degree of convexity, it is possible to search for a solution in its vicinity, thereby shortening the time required to search for a solution. As a result, delays in control, etc. can be suppressed.

[0033] Second Embodiment A second embodiment of the present disclosure will be described with reference to FIG. 7 . FIG. 7 is a diagram for explaining the configuration of an information processing device. As shown in FIG. 7 , the information processing device 10 of this embodiment includes a first optimization unit 13, a second optimization unit 14, and a learning model storage unit 17 in addition to the configuration of the information processing device of the first embodiment described above. The functions of the first optimization unit 13 and the second optimization unit 14 can be realized by a computing device executing a program for realizing each function stored in a storage device. The learning model storage unit 17 is also configured by a storage device. Below, each component will be described, mainly focusing on the components that differ from those of the first embodiment.

[0034] The learning model storage unit 17 stores a learning model f(x) generated by learning using the learning data stored in the learning data storage unit 16. Note that this learning model (x) is assumed to have been learned without taking into consideration the above-mentioned convexity parameter α.

[0035] The learning data storage unit 16 and the learning unit 11 have the same configuration as in the first embodiment. That is, the learning unit 11 uses the learning data stored in the learning data storage unit 16 to change the convexity parameter α and calculates a plurality of approximation functions f α Generate (x).

[0036] The first optimization unit 13 finds a constraint function g(f(x)) based on the learning model f(x) and the constraints set in advance for the learning model f(x), and finds a solution that minimizes this g(f(x)). For example, in a robot motion plan, if the objective variable of the learning model f(x) is feasibility, the constraint is assumed to be "70%". In this case, the constraint can be expressed as f(x) ≧ 0.7, and therefore the constraint function can be approximately expressed as g(f(x)) = 0.7 - f(x) ≦ 0. The first optimization unit 13 finds a solution (x) that minimizes this constraint function g. １ ＊ , x ２ ＊ , ..., x ｍ ＊ However, the first optimization unit 13 may use any method to find the constraint function based on the learning model f(x) and its constraints, and may use any method to find the solution of the constraint function.

[0037] The second optimization unit 14 has almost the same configuration as the optimization unit 12 of the first embodiment. The second optimization unit 14 optimizes the solution (x １ ＊ , x ２ ＊ , ..., x ｍ ＊ ) as an initial solution of the learning model f(x). That is, the second optimization unit 14 searches for a solution using the above-mentioned approximate function f α (x), the solution of the constraint function (x １ ＊ , x ２ ＊ , ..., x ｍ ＊ ) is given as an initial solution. At this time, the second optimization unit 14 sequentially optimizes the corresponding approximation functions fα (x) is used to find the solution and perform optimization. At this time, the approximate function f with the lowest value of α is α (x) or randomly select any approximation function f α (x) as the initial solution. After that, as in the first embodiment, the approximation function f α The solution obtained in (x) is applied to another approximation function f α (x) is given as the initial solution and the solution search is repeated to find the optimal solution.

[0038] At this time, the second optimization unit 14 calculates the approximate function f α When performing optimization using (x), a search for a solution may be performed using an annealing optimization method, in which the temperature parameter may be changed from low to high.

[0039] As described above, in this embodiment, a solution is first obtained based on the constraints, and the obtained solution is given as an initial solution for one of multiple approximate functions with different degrees of convexity. This makes it possible to shorten the search time for the optimal solution and suppress delays in control, etc.

[0040] Third Embodiment Next, a third embodiment of the present disclosure will be described with reference to Fig. 8 to Fig. 9. Fig. 8 to Fig. 9 are block diagrams showing the configuration of an information processing device in the third embodiment. Note that this embodiment shows an outline of the configuration of the information processing device described in the above-mentioned embodiments.

[0041] First, the hardware configuration of the information processing device 100 in this embodiment will be described with reference to Fig. 8. The information processing device 100 is configured as a general information processing device, and is equipped with the following hardware configuration, for example: CPU (Central Processing Unit) 101 (arithmetic unit); ROM (Read Only Memory) 102 (storage device); RAM (Random Access Memory) 103 (storage device); programs 104 loaded into RAM 103; storage device 105 storing programs 104; drive device 106 for reading and writing data from and to a storage medium 110 external to the information processing device; communication interface 107 for connecting to a communication network 111 external to the information processing device; input / output interface 108 for inputting and outputting data; and bus 109 for connecting the various components.

[0042] 8 shows an example of the hardware configuration of the information processing device 100, and the hardware configuration of the information processing device is not limited to the above-described case. For example, the information processing device may be configured with only a part of the above-described configuration, such as excluding the drive device 106. Furthermore, the information processing device may use a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point Number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof, instead of the above-described CPU.

[0043] The information processing device 100 can be equipped with the learning unit 121 and calculation unit 122 shown in FIG. 9 by having the CPU 101 acquire and execute the program group 104. The program group 104 is stored in advance in the storage device 105 or the ROM 102, for example, and is loaded into the RAM 103 and executed by the CPU 101 as needed. The program group 104 may be supplied to the CPU 101 via the communication network 111, or may be stored in advance in the storage medium 110, and the drive device 106 may read out the program and supply it to the CPU 101. However, the learning unit 121 and calculation unit 122 described above may be constructed using dedicated electronic circuits for realizing such means.

[0044] When learning and generating an approximation function that approximates the distribution of a dataset including explanatory variables and a target variable, the learning unit 121 learns and generates a plurality of approximation functions from the dataset using a loss function based on the degree of convexity set for the approximation function. At this time, the learning unit 121 learns and generates each of the approximation functions according to the degree of convexity using a different loss function depending on the degree of convexity.

[0045] The calculation unit 122 obtains a solution for the objective variable based on a plurality of approximation functions. At this time, the calculation unit 122 obtains a solution using approximation functions corresponding to the degrees of convexity in order from high to low.

[0046] With the above configuration, the present disclosure learns multiple approximation functions with different degrees of convexity and uses these approximation functions to find a solution. This makes it easier to find a solution for approximation functions with a high degree of convexity. Furthermore, because solutions are searched for sequentially from approximation functions with a high degree of convexity to approximation functions with a low degree of convexity, an optimal solution can be easily found in a short time. As a result, delays in control, etc. can be reduced.

[0047] The above-described program can be stored and supplied to a computer using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). The program may also be supplied to a computer by various types of transitory computer-readable media. Examples of transitory computer-readable media include electrical signals, optical signals, and electromagnetic waves. The transitory computer-readable media can be supplied to a computer via wired communication paths such as electric wires and optical fibers, or via wireless communication paths.

[0048] Although the present disclosure has been described above with reference to the above-described embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, at least one or more of the functions of the learning unit 121 and the calculation unit 122 described above may be executed by an information processing device installed and connected anywhere on a network, that is, may be executed by so-called cloud computing.

[0049] <Supplementary Notes> Some or all of the above embodiments may be described as in the following supplementary notes. Below, an outline of the configurations of an information processing device, an information processing method, and a program according to the present disclosure will be described. However, the present disclosure is not limited to the following configurations. (Supplementary Note 1) An information processing device comprising: a learning unit that, when learning and generating approximation functions that approximate the distribution of a dataset including explanatory variables and a dependent variable, learns and generates a plurality of the approximation functions from the dataset using a loss function based on a degree of convexity set for the approximation functions; and a calculation unit that obtains a solution for the dependent variable based on the plurality of approximation functions. (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the learning unit learns and generates each of the approximation functions according to the degree of convexity using the loss function whose value differs depending on the degree of convexity. (Supplementary Note 3) The information processing device according to Supplementary Note 2, wherein the learning unit learns and generates each of the approximation functions according to the degree of convexity using the loss function whose value increases as the degree of convexity increases. (Supplementary Note 4) The information processing device according to Supplementary Note 1, wherein the calculation unit calculates the solution using the approximation functions corresponding to the degrees of convexity in order from high to low. (Supplementary Note 5) The information processing device according to Supplementary Note 4, wherein the calculation unit searches for a solution by using a solution calculated from an approximation function in an earlier order as an initial solution for the approximation function in a later order, in the approximation functions for which the order of calculating the solution varies. (Supplementary Note 6) The information processing device according to Supplementary Note 1, wherein the calculation unit calculates a solution of a constraint function generated based on a learning model based on the dataset and constraints set on the learning model, and calculates the solution based on the approximation function, using the solution of the constraint function as an initial solution for the approximation function. (Supplementary Note 7) The information processing device according to Supplementary Note 6, wherein the calculation unit determines the solution based on one of the approximation functions using a solution of the constraint function as an initial solution of the approximation function, and determines the solution using the approximation functions corresponding to the degrees of convexity in order from higher to lower degrees of convexity.(Supplementary Note 8) The information processing device according to Supplementary Note 1, having a function of controlling a control target in accordance with the obtained solution. (Supplementary Note 9) An information processing method, comprising: when learning and generating an approximation function that approximates a distribution of a dataset including explanatory variables and a response variable, learning and generating a plurality of the approximation functions from the dataset using a loss function based on a degree of convexity set for the approximation function, and obtaining a solution for the response variable based on the plurality of the approximation functions. (Supplementary Note 10) The information processing method according to Supplementary Note 9, wherein the approximation functions corresponding to the degrees of convexity are learned and generated using the loss function having a different value depending on the degree of convexity. (Supplementary Note 11) The information processing method according to Supplementary Note 10, wherein the approximation functions corresponding to the degrees of convexity are learned and generated using the loss function whose value increases as the degree of convexity increases. (Supplementary Note 12) The information processing method according to Supplementary Note 9, wherein the solution is obtained using the approximation functions corresponding to the degrees of convexity in order from high to low. (Supplementary Note 13) The information processing method according to Supplementary Note 12, wherein, for the approximation functions whose order of solution determination changes, a solution is searched for by using a solution obtained from an earlier approximation function as an initial solution for the later approximation function. (Supplementary Note 14) The information processing method according to Supplementary Note 9, wherein a solution of a constraint function is determined based on a learning model based on the dataset and constraints set on the learning model, and the solution is determined based on the approximation function with the solution of the constraint function as an initial solution for the approximation function. (Supplementary Note 15) The information processing method according to Supplementary Note 14, wherein the solution is obtained based on one of the approximation functions using the solution of the constraint function as an initial solution for the approximation function, and the solution is obtained using the approximation functions corresponding to the degrees of convexity in order from high to low.(Supplementary Note 16) A computer-readable storage medium storing a program for executing, on a computer, a process of: when learning and generating an approximation function that approximates a distribution of a dataset including explanatory variables and a response variable, learning and generating a plurality of the approximation functions from the dataset using a loss function based on a degree of convexity set for the approximation function; and finding a solution for the response variable based on the plurality of the approximation functions.

[0050] REFERENCE SIGNS LIST 10 Information processing device 11 Learning unit 12 Optimization unit 13 First optimization unit 14 Second optimization unit 16 Learning data storage unit 17 Learning model storage unit 100 Information processing device 101 CPU 102 ROM 103 RAM 104 Program group 105 Storage device 106 Drive device 107 Communication interface 108 Input / output interface 109 Bus 110 Storage medium 111 Communication network 121 Learning unit 122 Calculation unit

Claims

1. a learning unit that, when learning and generating approximation functions that approximate the distribution of a dataset including explanatory variables and a target variable, learns and generates a plurality of the approximation functions from the dataset using a loss function based on a convexity degree set for the approximation functions; a calculation unit that calculates a solution for the response variable based on a plurality of the approximation functions; An information processing device comprising:

2. 2. The information processing device according to claim 1, the learning unit uses the loss function having a different value depending on the degree of convexity to learn and generate the approximation function depending on the degree of convexity, Information processing device.

3. 3. The information processing device according to claim 2, the learning unit learns and generates the approximation functions according to the degrees of convexity using the loss function, which has a value that increases as the degree of convexity increases; Information processing device.

4. 2. The information processing device according to claim 1, the calculation unit calculates the solution by using the approximation function corresponding to the degree of convexity in order from high to low. Information processing device.

5. 5. The information processing device according to claim 4, the calculation unit searches for a solution by using a solution obtained from an earlier approximation function as an initial solution for a later approximation function, when the order of obtaining the solution varies. Information processing device.

6. 2. The information processing device according to claim 1, the calculation unit obtains a solution of a constraint function generated based on a learning model based on the data set and constraints set in the learning model, and obtains the solution based on the approximation function using the solution of the constraint function as an initial solution of the approximation function; Information processing device.

7. 7. The information processing device according to claim 6, the calculation unit determines the solution based on one of the approximation functions using a solution of the constraint function as an initial solution of the approximation function, and determines the solution using the approximation functions corresponding to the degrees of convexity in order from higher to lower degrees of convexity. Information processing device.

8. 2. The information processing device according to claim 1, and controlling the control target in accordance with the obtained solution. Information processing device.

9. An information processing device, When learning and generating an approximation function that approximates a distribution of a dataset including explanatory variables and a target variable, a loss function based on a convexity degree set for the approximation function is used to learn and generate a plurality of the approximation functions from the dataset; determining a solution for the response variable based on a plurality of the approximation functions; Information processing methods.

10. When learning and generating an approximation function that approximates a distribution of a dataset including explanatory variables and a target variable, a loss function based on a convexity degree set for the approximation function is used to learn and generate a plurality of the approximation functions from the dataset; determining a solution for the response variable based on a plurality of the approximation functions; A program that executes processing on a computer.