Information processing apparatus, information processing method, and computer program

Abstract

To provide an information processing apparatus with which the composition of a material that satisfies target properties can be easily searched, compared to using Bayesian optimization or random search.SOLUTION: An information processing apparatus includes: a generation unit which generates a predetermined number of sets of proportion using Dirichlet distribution; and a prediction unit which predicts properties to be obtained from the sets of proportion generated by the generation unit. The generation unit generates a predetermined number of sets of proportion by providing, as a next parameter, to the Dirichlet distribution, values generated based on the predetermined number of proportion sets corresponding to the top-ranked properties predicted by the prediction unit, out of the proportion sets generated by providing an unbiased initial parameter to the Dirichlet distribution, and repeats generating, by the generation unit, sets of proportions using the Dirichlet distribution and predicting properties by the prediction unit, until the properties predicted by the prediction unit reach target property.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, an information processing method, and a computer program. [Background technology]

[0002] Bayesian optimization is generally used in composition search to find compositions of materials with excellent properties. For example, Patent Document 1 discloses techniques related to Bayesian optimization. When the search space is large, random search is sometimes used in composition search. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2023 / 084776 Summary of the Invention [Problem to be solved by the invention]

[0004] However, even if Bayesian optimization or random search is used in composition search to find a material composition with excellent properties, it may not be easy to find a composition that satisfies the target properties.

[0005] The present disclosure has been made in consideration of the above points, and aims to provide an information processing device, an information processing method, and a computer program that can easily search for a substance composition that satisfies target properties, compared to using Bayesian optimization or random search. [Means for solving the problem]

[0006] According to one aspect of the present disclosure, there is provided an information processing device comprising: a generation unit that generates a predetermined number of ratio sets using a Dirichlet distribution; and a prediction unit that predicts characteristics to be obtained from the ratio sets generated by the generation unit, wherein the generation unit generates the predetermined number of ratio sets by giving a value to the Dirichlet distribution as a next parameter, based on a predetermined number of ratio sets whose characteristics are highest among the ratio sets generated by giving unbiased initial parameters to the Dirichlet distribution, and repeats the generation of ratio sets using the Dirichlet distribution by the generation unit and the prediction of the characteristics by the prediction unit until the characteristics predicted by the prediction unit reach a target characteristic.

[0007] The generation unit may repeat the generation of a set of ratios using a Dirichlet distribution and the prediction unit may predict the characteristics a predetermined number of times, and if the best characteristic in the prediction result by the prediction unit no longer changes, the generation unit may stop generating the set of ratios.

[0008] The prediction unit may input the set of ratios generated by the generation unit into a trained model and obtain output from the trained model to predict characteristics to be obtained from the set of ratios.

[0009] The generation unit may use training data of the trained model as the initial parameters.

[0010] The prediction unit may input the set of ratios generated by the generation unit into a predetermined function and obtain an output from the function, thereby predicting characteristics to be obtained from the set of ratios.

[0011] The generator may use parameters created using a uniform distribution as the initial parameters.

[0012] According to another aspect of the present disclosure, there is provided an information processing method, the method comprising: a generation step in which a processor generates a predetermined number of ratio sets using a Dirichlet distribution; and a prediction step in which a characteristic to be obtained from the ratio sets generated in the generation step is predicted; the generation step generates the predetermined number of ratio sets by giving, to the Dirichlet distribution, as a next parameter, an average value of a predetermined number of ratio sets whose characteristics are highest among the ratio sets generated by giving unbiased initial parameters to the Dirichlet distribution; and the processor repeats the generation step of generating ratio sets using the Dirichlet distribution and the prediction of the characteristic in the prediction step until the characteristic predicted by the prediction step reaches a target characteristic.

[0013] According to another aspect of the present disclosure, there is provided a computer program for causing a computer to function as an information processing device, comprising: a generation unit that generates a predetermined number of ratio sets using a Dirichlet distribution; and a prediction unit that predicts characteristics to be obtained from the ratio sets generated by the generation unit, wherein the generation unit generates a predetermined number of ratio sets by applying an average value of a predetermined number of ratio sets whose characteristics are highest among the ratio sets generated by applying unbiased initial parameters to the Dirichlet distribution, and applies this average value to the Dirichlet distribution as a next parameter to generate a predetermined number of ratio sets, and repeats the generation of ratio sets using the Dirichlet distribution by the generation unit and the prediction of characteristics by the prediction unit until the characteristic predicted by the prediction unit reaches a target characteristic. [Effects of the Invention]

[0014] According to the present disclosure, it is possible to provide an information processing device, an information processing method, and a computer program that can easily search for a substance composition that satisfies target properties compared to using Bayesian optimization or random search. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a diagram illustrating an overview of an information processing device according to an embodiment of the disclosed technology. [Figure 2] FIG. 2 is a block diagram showing a hardware configuration of the information processing device. [Figure 3] FIG. 2 is a block diagram illustrating an example of a functional configuration of an information processing device. [Figure 4] 10 is a flowchart showing a flow of information processing by an information processing device. DETAILED DESCRIPTION OF THE INVENTION

[0016] An example of an embodiment of the present disclosure will be described below with reference to the drawings. The same reference numerals are used throughout the drawings to designate identical or equivalent components and parts. The dimensional proportions of the drawings are exaggerated for illustrative purposes and may differ from the actual proportions.

[0017] Fig. 1 is a diagram showing an overview of an information processing device according to this embodiment. The information processing device 10 shown in Fig. 1 is a device that executes a composition search to search for a substance composition having excellent properties, and is, for example, a personal computer, a server, or the like. As will be described later, the information processing device 10 is a device that executes a process of searching for a substance composition that satisfies target properties using a trained model 20.

[0018] When searching for a material composition with excellent properties, Bayesian optimization can be used if the search space is narrow, but it cannot be used if the search space is extremely wide. Furthermore, when searching for a material composition with excellent properties, if a random search is used, in which a set of ratios is randomly generated and given to a trained model, if a uniform distribution is used to generate a set of composition ratios for each material when mixing four types of materials, a composition set with uniform ratios such as [0.25, 0.25, 0.25, 0.25] is likely to be generated, but a composition set with uniform ratios does not necessarily have excellent properties. Furthermore, if a uniform distribution is used to generate a set of composition ratios, it is difficult to generate a biased composition set such as [0.84, 0.10, 0.04, 0.01], and compositions that have not been explored are likely to result.

[0019] In this embodiment, when searching for a material composition that satisfies target properties, the information processing device 10 uses a Dirichlet distribution to generate a predetermined number of sets of composition ratios (for example, 10,000 sets). By providing a parameter, the Dirichlet distribution can generate a set of random numbers that resemble the parameter. The total number of the set of random numbers generated is 1, and the information processing device 10 can treat the generated set of random numbers as composition ratios as they are.

[0020] First, the information processing device 10 generates a predetermined number of ratio sets by providing unbiased initial parameters to a Dirichlet distribution. Then, the information processing device 10 predicts characteristics using the generated ratio sets, and generates a predetermined number of ratio sets by providing values ​​generated based on a predetermined number of ratio sets with the highest predicted characteristics as the next parameters to the Dirichlet distribution. The information processing device 10 repeats generating ratio sets using the Dirichlet distribution and predicting characteristics using the generated ratio sets until the predicted characteristics reach the target characteristics.

[0021] The information processing device 10 can easily search for a material composition that satisfies the target characteristics compared to using Bayesian optimization or random search by repeatedly generating a set of ratios using a Dirichlet distribution and predicting the characteristics using the generated set of ratios until the predicted characteristics reach the target characteristics.

[0022] FIG. 2 is a block diagram showing the hardware configuration of the information processing device 10. As shown in FIG.

[0023] 2, the information processing device 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

[0024] The CPU 11 is a central processing unit that executes various programs and controls each component. That is, the CPU 11 reads the programs from the ROM 12 or storage 14 and executes the programs using the RAM 13 as a work area. The CPU 11 controls the above components and performs various arithmetic processing in accordance with the programs recorded in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores an information processing program for searching for a composition that satisfies target properties.

[0025] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured with a storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory, and stores various programs including the operating system and various data.

[0026] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to perform various inputs.

[0027] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may also function as the input unit 15 by adopting a touch panel system.

[0028] The communication interface 17 is an interface for communicating with other devices, and uses standards such as Ethernet (registered trademark), FDDI, and Wi-Fi (registered trademark).

[0029] When executing the above information processing program, the information processing device 10 uses the above hardware resources to realize various functions. The functional configuration realized by the information processing device 10 will be described.

[0030] FIG. 3 is a block diagram showing an example of the functional configuration of the information processing device 10. As shown in FIG.

[0031] 3, the information processing device 10 has, as functional components, a generation unit 101 and a prediction unit 102. Each functional component is realized by the CPU 11 reading and executing an information processing program stored in the ROM 12 or the storage 14.

[0032] The generation unit 101 generates a predetermined number (for example, 10,000 sets) of material ratio sets used for predicting properties in the prediction unit 102. The generation unit 101 generates the predetermined number (for example, 10,000 sets) of ratio sets using a Dirichlet distribution.

[0033] As a specific example, the generating unit 101 generates a set of ratios of each material when mixing PMMA (polymethyl methacrylate), PVDF (polyvinylidene fluoride), a compatibilizer, and an antioxidant to generate a polymer. The generating unit 101 also generates a set of ratios of each material when mixing propylene glycol, ethylene glycol, glycerin, and sorbitol to generate polyurethane.

[0034] The prediction unit 102 predicts the properties of the substance to be predicted using the set of ratios generated by the generation unit 101. In this embodiment, the prediction unit 102 inputs the set of ratios generated by the generation unit 101 into the trained model 20 and obtains output from the trained model 20. The output from the trained model 20 is a characteristic value that indicates the quality of the properties of the substance to be predicted. Any characteristic value can be used depending on the target substance. Examples of characteristic values ​​that can be used include mechanical strength, elongation, number of cracks, and degree of coloring.

[0035] In this embodiment, the generation unit 101 generates a set of ratios, and the prediction unit 102 predicts the properties of the substance to be predicted using the set of ratios, repeatedly until a composition that satisfies the target properties is found.

[0036] Specifically, the generation unit 101 first provides unbiased initial parameters to the Dirichlet distribution to generate a predetermined number of ratio sets. The generation unit 101 may use the training data of the trained model 20 as the initial parameters. For example, the generation unit 101 may extract a predetermined number of data sets that output excellent characteristics from the training data of the trained model 20 and use the average of the extracted data. Next, the prediction unit 102 inputs the ratio sets generated by the generation unit 101 into the trained model 20 and predicts characteristics by obtaining an output from the trained model 20.

[0037] In the Dirichlet distribution, the larger the sum of the parameters, the smaller the variance of the random numbers. The generator 101 may set the initial parameter sum to a small value to search a wide range, and may increase the parameter sum each time parameters are generated to search intensively near promising candidates.

[0038] Next, the generation unit 101 generates a predetermined number of ratio sets by providing the values ​​generated based on a predetermined number of ratio sets with the top characteristics predicted by the prediction unit 102 as the next parameters to the Dirichlet distribution. When the generation unit 101 generates 10,000 ratio sets using the initial parameters, the generation unit 101 generates the next parameters using, for example, the values ​​of the top 100 characteristic sets, and provides the generated parameters to the Dirichlet distribution to generate 10,000 ratio sets. The generation unit 101 uses, for example, the average value (arithmetic mean, weighted mean, harmonic mean, geometric mean, etc.) or median of the values ​​of each ratio as the value generated based on the predetermined number of ratio sets with the top characteristics. By the generation unit 101 generating the parameters in this way, the information processing device 10 can generate parameters to be provided to the Dirichlet distribution without arbitrariness.

[0039] The information processing device 10 repeats the generation of a set of ratios by the generation unit 101 and the prediction of the properties of the substance to be predicted using the set of ratios by the prediction unit 102 until a composition that satisfies the target properties is found, thereby making it easier to find a composition of a substance that satisfies the target properties compared to using Bayesian optimization or random search.

[0040] If the generation unit 101 generates a set of ratios and the prediction unit 102 predicts the properties of the substance to be predicted using the set of ratios a predetermined number of times (for example, five times) and the prediction result remains unchanged, i.e., if the best property value remains unchanged, the generation unit 101 may stop generating sets of ratios. The prediction unit 102 may use the set of ratios that derives the best property value as the composition of the substance that satisfies the target property.

[0041] The prediction unit 102 may impose a penalty on the neighborhood of the set of ratios that derives the best characteristic value, and execute a process to predict the characteristics of the substance to be predicted. In this case, the prediction unit 102 may execute the prediction process multiple times. The information processing device 10 can generate multiple sets of excellent ratios that are dissimilar to each other through this process. Note that the prediction unit 102 may use the distance from the set of excellent ratios (Euclidean distance, Manhattan distance, Mahalanobis distance, etc.) to determine the neighborhood. Furthermore, the prediction unit 102 may impose a penalty, such as subtracting infinity if the objective is to maximize the characteristic, or adding infinity if the objective is to minimize the characteristic.

[0042] Next, the operation of the information processing device 10 will be described.

[0043] 4 is a flowchart showing the flow of information processing by the information processing device 10. The CPU 11 reads out an information processing program from the ROM 12 or the storage 14, loads it into the RAM 13, and executes it, thereby performing information processing.

[0044] In step S101, the CPU 11 assigns initial parameters to the Dirichlet distribution to generate a predetermined number of ratio sets. In step S101, the CPU 11 first assigns unbiased initial parameters to the Dirichlet distribution to generate a predetermined number of ratio sets. The generation unit 101 may use the training data of the trained model 20 as the initial parameters.

[0045] Following step S101, in step S102, the CPU 11 provides the set of ratios generated in step S101 to the trained model 20 and obtains an output from the trained model 20, thereby predicting the properties of the substance to be predicted. The output is a characteristic value that indicates the quality of the properties of the substance to be predicted.

[0046] Following step S102, in step S103, the CPU 11 determines whether any of the characteristics predicted in step S102 satisfy the target characteristics.

[0047] If it is determined in step S103 that any of the characteristics predicted in step S102 satisfy the target characteristics (step S103; Yes), CPU 11 ends the series of processes. On the other hand, if it is determined in step S103 that any of the characteristics predicted in step S102 do not satisfy the target characteristics (step S103; No), following step S103, CPU 11 determines in step S104 whether characteristic prediction has been performed a predetermined number of times.

[0048] If the result of the determination in step S104 is that the characteristic prediction has been performed a predetermined number of times (step S104; Yes), the CPU 11 ends the series of processes. On the other hand, if the result of the determination in step S104 is that the characteristic prediction has not yet been performed a predetermined number of times (step S104; No), following step S104, the CPU 11 extracts in step S105 the sets of ratios that result in a predetermined number of top characteristics from the characteristics predicted in step S102.

[0049] Following step S105, in step S106, CPU 11 generates new parameters using the ratio sets extracted in step S105, and generates a predetermined number of new ratio sets by applying the generated parameters to a Dirichlet distribution. Specifically, CPU 11 sets the average value (arithmetic mean, weighted mean, harmonic mean, geometric mean, etc.) or median of the values ​​of each ratio in the ratio sets extracted in step S105 as the new parameters. After generating new parameters in step S106 and generating a predetermined number of new ratio sets, CPU 11 returns to the processing of step S102.

[0050] The information processing device 10 repeats the generation of a set of ratios and the prediction of the properties of the substance to be predicted using the set of ratios until a composition that satisfies the target properties is found, thereby making it easier to find a composition of a substance that satisfies the target properties compared to using Bayesian optimization or random search.

[0051] The present inventors attempted to discover the properties of polyurethane produced by preparing a polyol composition, mixing it with isocyanate, and molding it. In the case of a random search using a uniform distribution, the present inventors were unable to discover a composition with excellent properties even after generating 5 million ratio pairs, whereas, in the case of using a Dirichlet distribution as in the present embodiment, they were able to discover a composition with excellent properties after generating 40,000 ratio pairs.

[0052] A trained model generated by machine learning to predict properties from composition can answer the question of what properties a given composition will have, but it cannot answer the converse question of what composition will satisfy the target properties. In other words, it is necessary to use a trained model to exhaustively search for the ratio of a composition that satisfies the target properties.

[0053] However, using a random search to find the ratio of a composition that satisfies the target characteristics takes an enormous amount of time, and while it may be found quickly by chance, it may not be found even after a long time. In contrast, the information processing device 10 according to this embodiment can significantly reduce the time required to find a composition that satisfies the target characteristics compared to using Bayesian optimization or random search.

[0054] In the present embodiment, a trained model is used to predict the properties of the substance to be predicted, but the present disclosure is not limited to this example. Any function may be used instead of the trained model as long as the properties of the substance to be predicted can be output as a numerical value. When any function is used instead of the trained model, parameters created using a uniform distribution may be used as initial parameters for generating the ratio set.

[0055] Furthermore, in this embodiment, when there is an upper limit to the number of materials that can be used, the information processing device 10 may sample the likelihood of a material being selected from a Dirichlet distribution different from the Dirichlet distribution that generates the set of ratios.The information processing device 10 may then update the parameters of the Dirichlet distribution that samples the likelihood of a material being selected, based on the proportion of candidates with excellent properties that use a certain material.

[0056] Although the embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings, the technical scope of the present disclosure is not limited to such examples. The above-described embodiments are illustrative and do not limit the technical scope of the present disclosure. It is clear that a person skilled in the art of the present disclosure can conceive of various modifications or alterations within the scope of the technical idea described in the claims, and it is understood that these modifications or alterations also naturally fall within the technical scope of the present disclosure.

[0057] Furthermore, the effects described in the above embodiments are explanatory or exemplary and are not limited to those described in the above embodiments. In other words, the technology according to the present disclosure may achieve other effects that are obvious to a person skilled in the art of the present disclosure from the description in the above embodiments, in addition to or instead of the effects described in the above embodiments.

[0058] In the above embodiments, the information processing performed by the CPU after reading the software (program) may be performed by various processors other than the CPU. Examples of such processors include programmable logic devices (PLDs) whose circuit configuration can be changed after fabrication, such as field-programmable gate arrays (FPGAs), and dedicated electrical circuits, such as application-specific integrated circuits (ASICs), which are processors with circuit configurations specifically designed to perform specific processing. The information processing may be performed by one of these processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). The hardware structure of these processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor devices.

[0059] In addition, in each of the above embodiments, the information processing program is described as being pre-stored (installed) in a ROM or storage, but this is not limiting. The program may be provided in a form recorded on a non-transitory recording medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The program may also be downloaded from an external device via a network. [Explanation of symbols]

[0060] 10. Information processing equipment 101 Generation part 102 Prediction Department

Claims

1. a generation unit that generates a predetermined number of ratio sets using a Dirichlet distribution; a prediction unit that predicts characteristics obtained from the set of ratios generated by the generation unit; Equipped with the generation unit generates a predetermined number of ratio sets by giving a value generated based on a predetermined number of ratio sets having top characteristics predicted by the prediction unit, among the ratio sets generated by giving unbiased initial parameters to a Dirichlet distribution, to the Dirichlet distribution as a next parameter; an information processing apparatus that repeats the generation of a set of ratios using a Dirichlet distribution by the generation unit and the prediction of characteristics by the prediction unit until the characteristics predicted by the prediction unit reach target characteristics.

2. 2. The information processing device according to claim 1, wherein the generation unit generates a set of ratios using a Dirichlet distribution and the prediction unit predicts the characteristics a predetermined number of times, and when the best characteristic in the prediction result by the prediction unit no longer changes, the generation unit stops generating the set of ratios.

3. The information processing device according to claim 1 , wherein the prediction unit inputs the set of ratios generated by the generation unit into a trained model and obtains an output from the trained model to predict characteristics obtained from the set of ratios.

4. The information processing device according to claim 3 , wherein the generation unit uses training data of the trained model as the initial parameters.

5. The information processing device according to claim 1 , wherein the prediction unit predicts characteristics obtained from the set of ratios by inputting the set of ratios generated by the generation unit into a predetermined function and obtaining an output from the function.

6. The information processing device according to claim 1 , wherein the generating unit uses parameters generated using a uniform distribution as the initial parameters.

7. The processor: a generation step of generating a predetermined number of ratio sets using a Dirichlet distribution; a prediction step of predicting a characteristic obtained from the set of ratios generated in the generation step; Run the generating step generates a predetermined number of ratio sets by giving an average value of a predetermined number of ratio sets having top characteristics predicted by the predicting step to the Dirichlet distribution as a next parameter, among the ratio sets generated by giving unbiased initial parameters to the Dirichlet distribution; an information processing method in which the processor repeats generating a set of ratios using a Dirichlet distribution in the generating step and predicting the characteristic in the predicting step until the characteristic predicted in the predicting step reaches a target characteristic.

8. Computer, a generation unit that generates a predetermined number of ratio sets using a Dirichlet distribution; a prediction unit that predicts characteristics obtained from the set of ratios generated by the generation unit; Equipped with the generation unit generates a predetermined number of ratio sets by giving an average value of a predetermined number of ratio sets having top characteristics predicted by the prediction unit to the Dirichlet distribution as a next parameter, among the ratio sets generated by giving unbiased initial parameters to the Dirichlet distribution; A computer program for causing the computer to function as an information processing device, which repeats the generation of a set of ratios using a Dirichlet distribution by the generation unit and the prediction of characteristics by the prediction unit until the characteristics predicted by the prediction unit reach target characteristics.

Images