Method, information processing device, and program

Abstract

To provide a technology for more accurately predicting the probability of achieving a target according to the blending conditions of raw materials. [Solution] The method includes the steps of inputting input data including raw material blending conditions into a classification model and obtaining the probability that a material produced under those blending conditions will be classified into each of multiple evaluation results specified for a predetermined evaluation test. Each of the multiple pieces of performance data used to train the classification model associates the raw material blending conditions with the evaluation results of an evaluation test conducted on the material produced under those blending conditions. The method includes the steps of predicting, based on the multiple obtained probabilities, the target achievement probability that the material produced under the blending conditions specified in the input data will satisfy desired conditions, calculating, for each of the multiple pieces of performance data, a similarity between the blending conditions specified in the performance data and the blending conditions specified in the input data, and weighting the target achievement probability with weights corresponding to the multiple similarities.

Description

[Technical Field]

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

[0002] Japanese Patent Laid-Open Publication No. 2023-126824 (Patent Document 1) discloses a composition search method for more efficiently searching for a composition for obtaining a target physical property value.

[0003] The composition search method generates a prediction model by learning training data in which information about the composition of a material is used as an explanatory variable and the physical property values ​​of the material are used as target variables. The composition search method then inputs prediction data indicating the composition into the prediction model to calculate predicted values ​​of the physical properties. The composition search method then calculates the distance between the prediction data and the training data and weights the distance with a weight corresponding to the influence of each explanatory variable on the prediction. The composition search method then displays the relationship between the predicted value and the weighted distance. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2023-126824 Summary of the Invention [Problem to be solved by the invention]

[0005] Incidentally, there is a need to know the probability with which a material will achieve a target condition by changing the compounding conditions of the composition, raw materials, etc. In this regard, the composition search method disclosed in Patent Document 1 predicts the Young's modulus from the composition of the material (paragraph

[0101] ), but does not predict the probability of achieving the target.

[0006] The present disclosure has been made to solve the above-mentioned problems, and an object of one aspect is to provide a technology for more accurately predicting the probability of target achievement according to raw material blending conditions. [Means for solving the problem]

[0007] An example of the present disclosure provides a method executed by an information processing device. The method includes the steps of inputting input data including raw material blending conditions into a first classification model and obtaining a probability that a material produced under the blending conditions will be classified into each of multiple evaluation results specified for a predetermined first evaluation test. The first classification model is generated by a learning process using multiple first performance data. Each of the multiple first performance data associates the raw material blending conditions with the evaluation results of the first evaluation test conducted on the material produced under the blending conditions. The method includes the steps of predicting a first target achievement probability that a material produced under the blending conditions specified in the input data will satisfy a desired condition based on the multiple probabilities obtained in the obtaining step, calculating a similarity between the blending conditions specified in the first performance data and the blending conditions specified in the input data for each of the multiple first performance data, and weighting the first target achievement probability with a first weight corresponding to the multiple similarities calculated in the calculating step.

[0008] In one example of the present disclosure, the method further includes a step of weighting the first goal achievement probability weighted with the first weight with a second weight, and a step of outputting the sum of the first goal achievement probability weighted with the first weight and the second weight and the first goal achievement probability not weighted with the first weight and the second weight.

[0009] In one example of the present disclosure, the method further includes weighting the first goal achievement probability, which has not been weighted with the first weight, with a third weight, and the sum is a sum of the first goal achievement probability weighted with the first weight and the second weight and the first goal achievement probability weighted with the third weight.

[0010] In one example of the present disclosure, the method further includes the step of reducing the second weight as the number of the plurality of performance data increases.

[0011] In one example of the present disclosure, the method further includes a step of increasing the third weight as the number of the plurality of performance data increases.

[0012] In one example of the present disclosure, the sum of the second weight and the third weight is constant.

[0013] In one example of the present disclosure, the acquiring step inputs a plurality of pieces of input data to the first classification model, and the method includes a step of outputting input data from among the plurality of pieces of input data that maximizes the summation result.

[0014] In one example of the present disclosure, the first weight is calculated based on the similarity of the result data that is most similar to the input data among the plurality of result data.

[0015] In one example of the present disclosure, the blending conditions include at least one of the blending ratio of the raw materials and the manufacturing conditions when blending the raw materials.

[0016] In one example of the present disclosure, the material comprises a mucoadhesive material.

[0017] In one example of the present disclosure, the method further includes inputting the input data, including raw material blending conditions, into a second classification model and acquiring a probability that a material produced under the blending conditions will be classified into each of a plurality of evaluation results specified for a predetermined second evaluation test. The second classification model is generated by a learning process using a plurality of second performance data. Each of the plurality of second performance data associates raw material blending conditions with an evaluation result of the second evaluation test conducted on the material produced under the blending conditions. The method further includes predicting a second target achievement probability that the material produced under the blending conditions specified in the input data will satisfy a desired condition based on the plurality of probabilities acquired from the second classification model; calculating, for each of the plurality of second performance data, a similarity between the blending conditions specified in the second performance data and the blending conditions specified in the input data; weighting the second target achievement probability with a first weight corresponding to the plurality of similarities calculated for the input data; and integrating the first target achievement probability and the second target achievement probability.

[0018] Another example of the present disclosure provides an information processing device. The information processing device includes a control unit. The control unit inputs input data including raw material blending conditions into a first classification model and executes a process of acquiring a probability that a material produced under the blending conditions will be classified into each of a plurality of evaluation results specified for a predetermined first evaluation test. The first classification model is generated by a learning process using a plurality of first performance data. Each of the plurality of first performance data associates raw material blending conditions with an evaluation result of the first evaluation test conducted on the material produced under the blending conditions. The control unit further executes a process of predicting a first target achievement probability that a material produced under the blending conditions specified in the input data will satisfy a desired condition based on the plurality of probabilities acquired in the acquiring process; a process of calculating, for each of the plurality of first performance data, a similarity between the blending conditions specified in the first performance data and the blending conditions specified in the input data; and a process of weighting the first target achievement probability with a first weight corresponding to the plurality of similarities calculated in the calculating process.

[0019] Another example of the present disclosure provides a program executed by an information processing device. The program causes the information processing device to input input data including raw material blending conditions into a first classification model and acquire a probability that a material produced under the blending conditions will be classified into each of multiple evaluation results specified for a predetermined first evaluation test, the first classification model being generated by a learning process using multiple first performance data. Each of the multiple first performance data associates the raw material blending conditions with an evaluation result of the first evaluation test conducted on the material produced under the blending conditions. The program also causes the information processing device to predict a first target achievement probability that the material produced under the blending conditions specified in the input data will satisfy a desired condition, based on the multiple probabilities acquired in the acquiring process; calculate, for each of the multiple first performance data, a similarity between the blending conditions specified in the first performance data and the blending conditions specified in the input data; and weight the first target achievement probability with a first weight corresponding to the multiple similarities calculated in the calculating process.

[0020] The above and other objects, features, aspects and advantages of the present invention will become apparent from the following detailed description of the invention taken in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0021] [Figure 1] FIG. 2 is a diagram for explaining an outline of functions of an information processing device. [Figure 2] FIG. 2 illustrates an example of a hardware configuration of an information processing device. [Figure 3] FIG. 10 is a diagram illustrating an example of a training dataset. [Figure 4] FIG. 10 is a diagram illustrating an example of a prediction dataset. [Figure 5] FIG. 2 is a diagram illustrating an example of a functional configuration of an information processing apparatus. [Figure 6] FIG. 10 is a diagram conceptually illustrating a learning process performed by a learning unit. [Figure 7] FIG. 10 is a diagram conceptually illustrating a prediction process performed by an achievement probability calculation unit. [Figure 8] FIG. 10 is a diagram illustrating an example of an output result from an achievement probability calculation unit. [Figure 9] FIG. 10 is a diagram conceptually illustrating a prediction process performed by a distance weight calculation unit. [Figure 10] FIG. 10 is a diagram conceptually illustrating weighting processing by a correction unit. [Figure 11] 10 is a flowchart showing the flow of a learning process. [Figure 12] 10 is a flowchart showing the flow of a prediction process. [Figure 13] 10 is a flowchart showing the flow of a prediction process. [Figure 14] FIG. 10 is a diagram schematically illustrating a prediction process according to a second embodiment. [Figure 15] FIG. 10 is a diagram illustrating a training dataset according to a second embodiment. [Figure 16] FIG. 10 illustrates an example of a functional configuration of an information processing device according to a second embodiment. [Figure 17] FIG. 10 is a diagram conceptually illustrating a learning process according to a second embodiment. [Figure 18] FIG. 10 is a diagram conceptually illustrating a prediction process according to a second embodiment. [Figure 19] FIG. 10 is a diagram conceptually illustrating weighting processing according to the second embodiment. [Figure 20] FIG. 10 is a diagram conceptually illustrating integration processing according to the second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0022] Hereinafter, each embodiment according to the present invention will be described with reference to the drawings. In the following description, the same parts and components are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated. Note that each embodiment and each modified example described below may be selectively combined as appropriate.

[0023] [First Embodiment] <A. Overview> FIG. 1 is a diagram for explaining an overview of the functions of the information processing apparatus 100 according to the present embodiment.

[0024] The information processing apparatus 100 is, for example, a desktop PC (Personal Computer), a notebook PC, a tablet terminal, a smartphone, or other computer.

[0025] The information processing apparatus 100 has a function of predicting the target achievement probability that the material manufactured under the blending conditions satisfies the desired target conditions from the input data for prediction including the blending conditions of the raw materials. Hereinafter, an overview of the function of predicting the target achievement probability will be described.

[0026] The prediction of the target achievement probability is realized using the classification model 126. The classification model 126 has been generated in advance by a learning process using the learning dataset 122. The learning dataset 122 is composed of a plurality of performance data 123.

[0027] Each of the plurality of performance data 123 associates the blending conditions of the raw materials with the evaluation results of a predetermined evaluation test performed on the material manufactured under the blending conditions. In other words, in each of the plurality of performance data 123, the blending conditions of the raw materials are defined as explanatory variables, and the evaluation results are defined as target variables (labels). The evaluation result is indicated by, for example, an evaluation rank given in a qualitative evaluation test. A qualitative evaluation test is a test in which an evaluation scale that is difficult to be quantified by a device is artificially evaluated according to a predetermined standard. Specific examples of the qualitative evaluation test will be described later.

[0028] In step S1, the information processing apparatus 100 inputs the input data 125 including the blending conditions of the raw materials into the classification model 126. As a result, the classification model 126 outputs the probability (hereinafter, also referred to as "classification probability") that the material manufactured under the blending conditions defined in the input data 125 is classified into each evaluation result in the qualitative evaluation test.

[0029] Next, in step S2, the information processing device 100 predicts the target achievement probability that the material manufactured under the blending conditions specified in the input data 125 will satisfy the desired target conditions, based on the classification probabilities acquired in step S1. As an example, the information processing device 100 sums the classification probabilities for each evaluation rank belonging to the target conditions, and predicts the sum as the target achievement probability.

[0030] Next, in step S3, the information processing device 100 calculates, for each of the multiple performance data 123, the similarity between the blending conditions defined in the performance data 123 and the blending conditions defined in the input data 125. The similarity indicates how similar the input data 125 for prediction is to the learning dataset 122.

[0031] Any algorithm may be used to calculate the similarity, such as kernel density estimation (KDE), mean squared error (MSE), sum of squared difference (SSD), sum of absolute difference (SAD), normalized cross-correlation (NCC), or zero-mean normalized cross-correlation (ZNCC).

[0032] The magnitude of the calculated similarity varies depending on the algorithm used. That is, the calculated similarity value may be larger as the performance data 123 and the input data 125 are more similar to each other, or may be smaller as the performance data 123 and the input data 125 are more similar to each other.

[0033] Hereinafter, for ease of understanding, the similarity will be expressed as "distance." That is, a short distance between the performance data 123 and the input data 125 indicates that the performance data 123 and the input data 125 are similar. On the other hand, a long distance between the performance data 123 and the input data 125 indicates that the performance data 123 and the input data 125 are dissimilar.

[0034] Next, in step S4, the information processing device 100 weights the target achievement probability predicted in step S2 with weights (first weights) corresponding to the distances calculated in step S3. In this way, the information processing device 100 corrects the target achievement probability according to the distance from the performance data 123 used for learning. As a result, the information processing device 100 can more accurately predict the target achievement probability according to the blending conditions of the ingredients.

[0035] Typically, the information processing device 100 calculates the weighted target achievement probability for a plurality of input data 125. Then, the information processing device 100 outputs the input data 125 with the highest weighted target achievement probability as a prescription candidate. Alternatively, the information processing device 100 outputs the input data 125 with the weighted target achievement probability that falls within a predetermined top rank as a prescription candidate. This allows the evaluator to efficiently search for raw material blending conditions for a material that satisfies the desired target conditions.

[0036] Preferably, the weight according to the distance is calculated using the distance (hereinafter also referred to as the "minimum distance") relating to the actual data 123 that is most similar to the input data 125 among the actual data 123 included in the learning dataset 122.

[0037] In a certain situation, the information processing apparatus 100 corrects such that the smaller the value of the minimum distance, the smaller the target achievement probability. In other words, the information processing apparatus 100 corrects such that the larger the value of the minimum distance, the larger the target achievement probability. As a result, when there is no performance data 123 similar to the input data 125, the target achievement probability becomes larger. Thereby, it becomes easier for the input data 125 for which the accuracy of the prediction result cannot be compensated to remain as a prescription candidate, and the optimal prescription can be entrusted to the evaluator.

[0038] In another situation, the information processing apparatus 100 corrects such that the smaller the value of the minimum distance, the larger the target achievement probability. In other words, the information processing apparatus 100 corrects such that the larger the value of the minimum distance, the smaller the target achievement probability. As a result, when there is performance data 123 similar to the input data 125, the target achievement probability becomes larger. Thereby, it becomes easier for the input data 125 similar to the performance data 123 to remain as a prescription candidate.

[0039] <B. Hardware Configuration of Information Processing Apparatus 100> Next, referring to FIG. 2, the hardware configuration of the information processing apparatus 100 will be described. FIG. 2 is a diagram showing an example of the hardware configuration of the information processing apparatus 100.

[0040] The information processing apparatus 100 includes a control device 101 (control unit), a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, a communication interface 104, a display interface 105, an input interface 107, and an auxiliary storage device 120. These components are connected to a bus 115.

[0041] The control device 101 is configured, for example, by at least one integrated circuit. The integrated circuit may be configured, for example, by at least one central processing unit (CPU), at least one graphics processing unit (GPU), at least one application specific integrated circuit (ASIC), at least one field programmable gate array (FPGA), or a combination thereof.

[0042] The control device 101 controls the operation of the information processing device 100 by executing various programs. Upon receiving an execution command for one of the programs, the control device 101 reads the program to be executed from the auxiliary storage device 120 or the ROM 102 into the RAM 103. The RAM 103 functions as a working memory and temporarily stores various data required for executing the program.

[0043] A LAN (Local Area Network), an antenna, etc. are connected to the communication interface 104. The information processing device 100 exchanges data with external devices via the communication interface 104. The external devices include, for example, servers.

[0044] A display device 106 is connected to the display interface 105. The display interface 105 sends an image signal for displaying an image to the display device 106 in accordance with a command from the control device 101 or the like. The display device 106 is, for example, a liquid crystal display, an organic EL (Electro Luminescence) display, or other displays. The display device 106 may be configured integrally with the information processing device 100 or may be configured separately from the information processing device 100.

[0045] An input device 108 is connected to the input interface 107. The input device 108 is, for example, a mouse, a keyboard, a touch panel, or any other device capable of receiving user operations. The input device 108 may be configured integrally with the information processing device 100, or may be configured separately from the information processing device 100.

[0046] The auxiliary storage device 120 is a storage medium such as a hard disk or a flash memory. The auxiliary storage device 120 stores, for example, the above-mentioned training dataset 122, the above-mentioned prediction dataset 124 including the above-mentioned input data 125, the above-mentioned classification model 126, the training program 128, and the prediction program 130. These may be stored in a storage location other than the auxiliary storage device 120, such as a storage area of ​​the control device 101 (for example, a cache memory), the ROM 102, the RAM 103, an external device (for example, a server), or the like.

[0047] The learning program 128 is a program for generating the classification model 126 using the learning dataset 122. The learning program 128 may be provided not as a standalone program but as part of an arbitrary program. In this case, the learning process by the learning program 128 is realized in cooperation with the arbitrary program. Even a program that does not include some of these modules does not deviate from the spirit of the learning program 128 according to this embodiment. Furthermore, some or all of the functions provided by the learning program 128 may be realized by dedicated hardware. Furthermore, the information processing device 100 may be configured in the form of a so-called cloud service in which at least one server executes part of the processing of the learning program 128.

[0048] The prediction program 130 is a program for realizing the processing of the above steps S1 to S4. The prediction program 130 may be provided by being incorporated into a part of an arbitrary program instead of being a single program. In this case, the learning process by the prediction program 130 is realized in cooperation with an arbitrary program. Even a program that does not include such a part of the module does not deviate from the gist of the prediction program 130 according to the present embodiment. Furthermore, part or all of the functions provided by the prediction program 130 may be realized by dedicated hardware. Furthermore, the information processing apparatus 100 may be configured in a form such as a so-called cloud service in which at least one server executes a part of the processing of the prediction program 130.

[0049] <C. Learning dataset 122> Next, referring to FIG. 3, the learning dataset 122 shown in FIGS. 1 and 2 will be described. FIG. 3 is a diagram showing an example of the learning dataset 122.

[0050] The learning dataset 122 includes a plurality of performance data 123. The number of performance data 123 included in the learning dataset 122 is arbitrary. As an example, the number of performance data 123 is several to tens of thousands.

[0051] Each of the performance data 123 has a data ID (Identification) defined. The data ID is an identifier for uniquely identifying the performance data 123. The data ID is defined by, for example, a combination of characters, numerical values, and symbols.

[0052] Also, each of the performance data 123 associates an explanatory variable and an objective variable. The value of the explanatory variable defined in the performance data 123 and the value of the objective variable defined in the performance data 123 are measured values.

[0053] The explanatory variables defined in the performance data 123 are, for example, the blending conditions of the raw materials. The blending conditions include, for example, at least one of the blending ratio of each type of raw material and the manufacturing conditions when manufacturing the material from each raw material. Examples of the manufacturing conditions include the temperature during manufacturing, the pressure applied during manufacturing, and the manufacturing time.

[0054] The objective variables defined in the performance data 123 include evaluation ranks (evaluation results) assigned in qualitative evaluation tests. As described above, qualitative evaluation tests are tests that manually evaluate evaluation scales that are difficult to quantify using equipment according to predetermined criteria. For example, an evaluator visually checks a material and assigns an evaluation rank related to the physical properties of the material as a label according to predetermined evaluation criteria. The evaluation rank may be represented by two values, "OK" and "NG," or by three or more values.

[0055] Materials to be evaluated include, for example, adhesive materials and inks. "Adhesive materials" refers to materials that have the function of joining objects together. Examples of adhesive materials include adhesive tapes and adhesives.

[0056] An example of a qualitative evaluation test is a test for evaluating the adhesiveness of cellophane tape. In this evaluation test, whether or not a surface coating, paint film, print, etc. peels off when an adhesive tape attached to an object is peeled off is evaluated. The evaluator evaluates the adhesiveness of cellophane tape, for example, using a rating scale of "1" to "5." A rating scale of "5" indicates "no peeling." A rating scale of "4" indicates "slight peeling." A rating scale of "3" indicates "partial peeling." A rating scale of "2" indicates "considerable peeling." A rating scale of "1" indicates "almost complete peeling."

[0057] Another example of a qualitative evaluation test is a test for evaluating offset. In this evaluation test, it is evaluated whether ink applied to a sheet-like object or an adhesive material attached to a sheet-like object will penetrate to the back side or be transferred to another object. The evaluator evaluates the offset using an evaluation rank of "1" to "5," for example. As an example, an evaluation rank of "5" indicates "no offset." An evaluation rank of "4" indicates "almost no offset." An evaluation rank of "3" indicates "slight offset." An evaluation rank of "2" indicates "clear offset." An evaluation rank of "1" indicates "significant offset."

[0058] Another example of a qualitative evaluation test is an appearance evaluation test. Examples of appearance evaluation tests include a test to evaluate the presence or absence of scratches, a test to evaluate the presence or absence of cracks, a test to evaluate the presence or absence of foreign matter, a test to evaluate color, and a test to evaluate gloss. In an appearance evaluation test, an evaluator assigns an evaluation rank based on the appearance of the material.

[0059] The qualitative evaluation test is not limited to the visual evaluation test described above, but may be an evaluation test in which there is a maximum value in the evaluation criteria. Examples of such evaluation tests include a friction resistance test, a peel strength test, and a strength test of a laminating adhesive.

[0060] In the abrasion resistance test, an evaluator rubs a sample (such as ink) applied to a printed matter under certain conditions to check for changes in the surface state of the sample. Then, the evaluator records the number of rubbing times when the surface deteriorates as the evaluation value. The upper limit of the number of rubbing times is, for example, 200 times. That is, if the surface state does not change even when the evaluator rubs 200 times, it is considered qualified. In this case, the evaluation values of the qualified products are all the same and serve as qualitative evaluation indicators. The evaluator evaluates the abrasion resistance, for example, using either evaluation rank "0" or "1". More specifically, when the sample does not deteriorate even when the number of rubbing times reaches the upper limit, the evaluator assigns "1", indicating qualification, as the evaluation rank. On the other hand, when the sample deteriorates before the number of rubbing times reaches the upper limit, the evaluator assigns "0", indicating non - qualification, as the evaluation rank.

[0061] In the peel strength test, for example, the adhesive strength of a sample in which a base material such as a film is bonded with an adhesive material is evaluated. The evaluator evaluates the strength required for peeling the adhesive material and the base material as the adhesive strength. The adhesive strength is measured with a measuring instrument. The evaluator evaluates the peel strength, for example, using either evaluation rank "0" or "1". More specifically, when the measured adhesive strength reaches the pass criterion, the evaluator assigns "1", indicating qualification, as the evaluation rank. On the other hand, when the base material peels off before the adhesive strength reaches the upper limit, the evaluator assigns "0", indicating non - qualification, as the evaluation rank.

[0062] <D. Prediction Dataset 124> Next, referring to FIG. 4, the prediction dataset 124 shown in FIG. we will be described. FIG. 4 is a diagram showing an example of the prediction dataset 124.

[0063] The prediction dataset 124 includes a plurality of input data 125. The number of input data 125 included in the prediction dataset 124 is arbitrary. As an example, the number of input data 125 is from several to tens of thousands.

[0064] Each of the input data 125 has a data ID defined. The data ID is an identifier for uniquely identifying the input data 125. The data ID is defined, for example, as a combination of characters, numerical values, and symbols.

[0065] Also, each of the input data 125 defines the blending conditions of the raw materials as explanatory variables. The types of explanatory variables defined in the input data 125 are the same as the types of explanatory variables defined in the above-mentioned performance data 123 (see Figure 3). More specifically, the blending conditions defined in the input data 125 include, for example, at least one of the blending ratios of each raw material and the manufacturing conditions when manufacturing the material from each raw material. Examples of manufacturing conditions include the temperature during manufacturing, the pressure applied during manufacturing, and the manufacturing time.

[0066] The blending ratio defined for each of the input data 125 may be randomly determined within the range specified by the evaluator, or may be arbitrarily specified by the evaluator.

[0067] <E. Functional Configuration of Information Processing Apparatus 100> Next, referring to FIGS. 5 to 10, the functional configuration of the information processing apparatus 100 will be described. FIG. 5 is a diagram showing an example of the functional configuration of the information processing apparatus 100.

[0068] As shown in FIG. 5, the information processing apparatus 100 includes, as functional configurations, a learning unit 152, an achievement probability calculation unit 154, a distance weight calculation unit 156, a correction unit 158, and an output unit 160. Hereinafter, these functional configurations will be described in order.

[0069] Note that it is not necessary for all of the learning unit 152, the achievement probability calculation unit 154, the distance weight calculation unit 156, the correction unit 158, and the output unit 160 to be implemented in the information processing apparatus 100. Some of the functional configurations may be implemented in the information processing apparatus 100, and the remaining functional configurations may be implemented in another computer such as a server.

[0070] As an example, the achievement probability calculation unit 154, the distance weight calculation unit 156, the correction unit 158, and the output unit 160 may be implemented in the information processing device 100, and the learning unit 152 may be implemented in another computer.

[0071] (E1. Learning Section 152) First, the function of the learning unit 152 shown in Fig. 5 will be described with reference to Fig. 6. Fig. 6 is a diagram conceptually showing the learning process performed by the learning unit 152.

[0072] The learning unit 152 executes a learning process using the above-described learning dataset 122 composed of a plurality of pieces of performance data 123, and generates a classification model 126. The machine learning algorithm employed is not particularly limited, and various machine learning algorithms such as a neural network such as deep learning, a support vector machine, or a decision tree system may be employed. The learning process using a neural network will be described below.

[0073] The classification model 126 is composed of an input layer X, a hidden layer H, and an output layer Y.

[0074] The input layer X is configured to receive input of explanatory variables defined in the performance data 123. As an example, the input layer X includes a unit group x1 and a unit group x2.

[0075] The unit group x1 is configured to receive explanatory variables (e.g., raw material blending ratios) defined in the performance data 123. The number of units constituting the unit group x1 is the same as the number of dimensions of the blending ratios. As an example, if the blending ratios are five-dimensional, the unit group x1 is composed of five units. Each unit constituting the unit group x1 outputs the input data to each unit in the first layer of the intermediate layer H.

[0076] The unit group x2 is configured to receive input of other explanatory variables (e.g., manufacturing conditions) defined in the classification model 126. The number of units constituting the unit group x2 is the same as the number of dimensions of the manufacturing conditions. As an example, if the manufacturing conditions are expressed in two dimensions, the unit group x2 is composed of two units. Each unit constituting the unit group x2 outputs the input data to each unit in the first layer of the intermediate layer H.

[0077] The middle layer H is composed of multiple layers. The number of layers in the middle layer H is arbitrary. In the example of FIG. 6, the middle layer H is composed of N layers (N is a natural number). Each layer in the middle layer H includes multiple units. In the example of FIG. 6, the first layer of the middle layer H is composed of unit h A1 ,h A2 The final layer of the intermediate layer H is made up of units h N1 ,h N2 It is composed of...

[0078] Each unit constituting each layer of the hidden layer H is connected to each unit in the previous layer and each unit in the next layer. Each unit in each layer receives each output value from each unit in the previous layer, multiplies each output value by a weight, accumulates the multiplication results, adds (or subtracts) a predetermined bias to (or from) the accumulation result, inputs the addition result (or subtraction result) to a predetermined function (for example, a sigmoid function), and outputs the output value of the function to each unit in the next layer.

[0079] The output layer Y outputs an evaluation result (i.e., classification probability) according to the explanatory variables input to the input layer X. As an example, the output layer Y is composed of units y1 to y3. Hereinafter, the units y1 to y3 will also be referred to as unit y.

[0080] Each unit y is connected to each unit h in the final layer of the hidden layer H. N1 ,h N2Each of the units y receives an output value from each unit in the final layer of the hidden layer H, multiplies each output value by a weight, accumulates the results of these multiplications, adds (or subtracts) a predetermined bias to (or from) the accumulated result, inputs the result of the addition (or subtraction) to a predetermined function (for example, a sigmoid function), and outputs the output result of the function as an output value.

[0081] The number of units constituting the output layer Y is determined according to the number of objective variables defined in the performance data 123. As described above, the objective variables are evaluation ranks given in a qualitative evaluation test, and the number of units constituting the output layer Y is, for example, the same as the number of evaluation ranks.

[0082] Unit y1 is configured to output a classification probability belonging to evaluation rank "1." Unit y2 is configured to output a classification probability belonging to evaluation rank "2." Unit y3 is configured to output a classification probability belonging to evaluation rank "3."

[0083] Note that the output value of the classification model 126 does not need to be a probability itself, but may be a value correlated with the probability. As an example, the output value of the classification model 126 may be a score correlated with the classification probability. In this case, the information processing device 100 normalizes the scores so that the sum of the scores becomes 1, and calculates the normalized score as the classification probability.

[0084] Furthermore, in the example of FIG. 6, one classification model 126 is configured to output a classification probability belonging to evaluation ranks "1" to "3", but one classification model 126 may output one probability. As an example, the first classification model 126 is configured to receive input of the blending conditions of the ingredients and output a probability belonging to evaluation rank "1". The second classification model 126 is configured to receive input of the blending conditions of the ingredients and output a probability belonging to evaluation rank "2". The third classification model 126 is configured to receive input of the blending conditions of the ingredients and output a probability belonging to evaluation rank "3".

[0085] Next, the update process of the internal parameters of the classification model 126 by the learning unit 152 will be described.

[0086] The learning unit 152 acquires the first piece of performance data 123 from the training dataset 122, and inputs the explanatory variables defined in the performance data 123 to the classification model 126. As a result, the classification model 126 outputs the probability of belonging to each evaluation rank. Next, the learning unit 152 compares the output prediction results "p1" to "p3" with the objective variable defined in the first piece of performance data 123. The objective variable is expressed, for example, as probabilities "YA" to "YC".

[0087] As an example, when the evaluation rank associated with the performance data 123 is "1", the value of the objective variable is (YA, YB, YC) = (1, 0, 0). When the evaluation rank associated with the performance data 123 is "2", the value of the objective variable is (YA, YB, YC) = (0, 1, 0). When the evaluation rank associated with the performance data 123 is "3", the value of the objective variable is (YA, YB, YC) = (0, 0, 1).

[0088] The learning unit 152 calculates the error "Z" between the output results "p1" to "p3" of the classification model 126 and the objective variables "YA" to "YC." As an example, the error "Z" is calculated based on the following formula (1).

[0089] Z={(p1-YA) 2 +(p2-YB) 2 +(p3-YC) 2} / 3···(1) Next, the learning unit 152 updates various parameters (for example, weights and biases) included in the classification model 126 so as to reduce the error "Z." The parameter update is realized, for example, by the backpropagation algorithm.

[0090] The learning unit 152 repeatedly updates the internal parameters of the classification model 126 for each piece of performance data 123 included in the learning dataset 122. As a result, the classification model 126 begins to output accurate prediction results as the learning progresses.

[0091] Note that the learning unit 152 does not need to use all of the performance data 123 included in the training data set 122 for the learning process, and may generate the classification model 126 using a portion of the performance data 123 included in the training data set 122. The remaining performance data 123 is used, for example, to evaluate the classification model 126.

[0092] (E2. Achievement Probability Calculation Unit 154) Next, the function of the achievement probability calculation unit 154 shown in Fig. 5 will be described with reference to Fig. 7. Fig. 7 is a diagram conceptually showing the prediction process by the achievement probability calculation unit 154.

[0093] The achievement probability calculation unit 154 receives input of target conditions 132 related to the physical properties of the material. The target conditions 132 are input, for example, to an input screen displayed on the above-mentioned display device 106. The evaluator inputs the target conditions 132 to the input screen using, for example, the above-mentioned input device 108.

[0094] The target condition 132 is a condition for specifying a range of evaluation ranks associated with the output layer of the classification model 126. In this case, the target condition 132 is defined by at least one threshold value.

[0095] Next, the achievement probability calculation unit 154 sequentially acquires the input data 125 included in the above-mentioned prediction data set 124 (see FIG. 4 ), and inputs the input data 125 to the classification model 126. As a result, the classification model 126 outputs the probability of being classified into each evaluation rank defined for the qualitative evaluation test.

[0096] When a numerical range equal to or less than a predetermined threshold is specified as the goal condition 132, the achievement probability calculation unit 154 sums up the classification probabilities for the evaluation ranks equal to or less than the predetermined threshold, and calculates the sum result as the goal achievement probability "P pred Here, "i" is the number of the input data 125 defined in the prediction dataset 124. As an example, when the input goal condition 132 indicates an evaluation rank of "2" or less, the achievement probability calculation unit 154 calculates the goal achievement probability "P (i)" as the sum of the classification probability "p1" for the evaluation rank "1" and the classification probability "p2" for the evaluation rank "2". pred (i)" is calculated as follows.

[0097] When a numerical range equal to or greater than a predetermined threshold is specified as the goal condition 132, the achievement probability calculation unit 154 sums up the classification probabilities for evaluation ranks equal to or greater than the threshold, and calculates the sum as the goal achievement probability "P pred As an example, when the input goal condition 132 indicates an evaluation rank of "2" or higher, the achievement probability calculation unit 154 calculates the goal achievement probability "P (i)" as the sum of the classification probability "p2" for the evaluation rank "2" and the classification probability "p3" for the evaluation rank "3." pred (i)" is calculated as follows.

[0098] When a numerical range equal to or greater than the first threshold and equal to or less than the second threshold is specified as the goal condition 132, the achievement probability calculation unit 154 sums up the classification probabilities for the evaluation ranks that are equal to or greater than the first threshold and equal to or less than the second threshold, and calculates the sum result as the goal achievement probability "P pred (i)" is calculated as follows.

[0099] 8 is a diagram showing an output result RS1 by the achievement probability calculation unit 154. As shown in FIG. 8, the output result RS1 is calculated for each piece of input data 125 input to the classification model 126. The output result RS1 is output to the correction unit 158.

[0100] (E3. Distance weight calculation unit 156) Next, the function of distance weight calculation unit 156 shown in Fig. 5 will be described with reference to Fig. 9. Fig. 9 is a diagram conceptually showing the prediction process by distance weight calculation unit 156.

[0101] The distance weight calculation unit 156 calculates, for each of the input data 125 to be predicted, the distance between the combination conditions defined in the input data 125 and the combination conditions defined in the performance data 123. The method for calculating the distance is as described above, and therefore the description thereof will not be repeated.

[0102] Next, the distance weight calculation unit 156 identifies the minimum distance between the actual data 123 and the input data 125 for each of the input data 125 to be predicted. Hereinafter, the minimum distance identified for each of the input data 125 is referred to as "d0(i)".

[0103] 9 shows the output result RS2 by the distance weight calculation unit 156. The output result RS2 shows the minimum distance "d0(i)" identified for each piece of input data 125 to be predicted.

[0104] Next, the distance weight calculation unit 156 normalizes the minimum distance "d0(i)" so that it is equal to or greater than 0 and equal to or less than 1. More specifically, the distance weight calculation unit 156 first calculates the minimum distance "d min Next, the distance weight calculation unit 156 determines the maximum distance "d max Next, the distance weight calculation unit 156 calculates the normalized distance "d(i)" based on the following equation (2).

[0105] d(i)=(d0(i)-d min ) / (d max -d min )···(2)

[0106] (E4. Correction unit 158) Next, the function of the correction unit 158 ​​shown in Fig. 5 will be described with reference to Fig. 10. Fig. 10 is a diagram conceptually showing the weighting process performed by the correction unit 158.

[0107] 10 shows the output result RS1 from the achievement probability calculation unit 154, the output result RS2 from the distance weight calculation unit 156, and the output result RS3 from the correction unit 158. The correction unit 158 ​​corrects the goal achievement probability "P" calculated by the achievement probability calculation unit 154 using a weight according to the distance "d(i)" calculated by the distance weight calculation unit 156. pred As an example, the correction unit 158 ​​corrects the weighted goal achievement probability "P calc (i)" is calculated.

[0108] P calc (i)=P pred (i)·d(i)···(3) In addition, the target achievement probability "P calc The calculation formula for "d(i)" is not limited to the above formula (3). As another example, the correction unit 158 ​​may calculate the goal achievement probability "P calc (i)" is further weighted with a weight "w2" (second weight). Then, the correction unit 158 ​​calculates the weighted goal achievement probability "P pred (i)” and the unweighted goal achievement probability “P pred (i)” and the sum of them is the target achievement probability “P calc (i)” can be used. In this case, the target achievement probability “P calc (i)" is calculated based on the following formula (4).

[0109] P calc (i)=P pred (i)+w2·P pred (i)·d(i)···(4) The weight "w2" shown in the above formula (4) is, for example, a value greater than or equal to 0 and less than or equal to 1. The weight "w2" may be arbitrarily set by the evaluator, or may be set in advance.

[0110] As described above, the correction unit 158 ​​calculates the goal achievement probability "P predThe correction unit 158 ​​may further weight "d(i)" with a weight "w2". Then, the correction unit 158 ​​calculates the goal achievement probability "P pred (i)" and the goal achievement probability "P pred (i)” and the sum of them is the target achievement probability “P calc By adjusting the weight "w2", the evaluator determines whether the distance "d(i)" is equal to the target achievement probability "P calc The degree of influence on (i) can be changed arbitrarily.

[0111] As another example, the correction unit 158 ​​calculates the weighted goal achievement probability "P calc (i)" may be calculated.

[0112] P calc (i)=w3·P pred (i)+w2·P pred (i)·d(i)···(5) The weight "w3" shown in the above formula (5) is, for example, a value greater than or equal to 0 and less than or equal to 1. The weight "w3" may be arbitrarily set by the evaluator, or may be set in advance.

[0113] As shown in the above formula (5), the correction unit 158 ​​calculates the goal achievement probability "P pred Then, the correction unit 158 ​​weights the target achievement probability "P pred (i)” and the goal achievement probability “P pred (i)” and the sum of them is the target achievement probability “P calc (i)" is output.

[0114] The weight "w2" shown in the above formula (4) or (5) may be automatically adjusted. In this case, the correction unit 158 ​​changes the weight "w2" based on the number of actual data 123 used during learning. As an example, the correction unit 158 ​​reduces the weight "w2" as the number of actual data 123 increases. In other words, the correction unit 158 ​​increases the weight "w2" as the number of actual data 123 decreases.

[0115] Furthermore, the weight "w3" shown in the above formula (5) may be automatically adjusted. In this case, the correction unit 158 ​​changes the weight "w3" based on the number of performance data 123 used during learning. As an example, the correction unit 158 ​​increases the weight "w3" as the number of performance data 123 increases. In other words, the correction unit 158 ​​decreases the weight "w3" as the number of performance data 123 decreases.

[0116] Furthermore, the sum of the weights "w2" and "w3" shown in the above formula (5) may be constant. calc The scale of (i) can be made constant. Typically, the sum of the weights "w2" and "w3" is 1. In this case, the weight "w3" follows the following formula (6).

[0117] w3=1-w2 (6)

[0118] (E5. Output unit 160) Continuing to refer to FIG. 10, the function of the output unit 160 shown in FIG. 5 will be described.

[0119] The output unit 160 outputs the target achievement probability "P calc Based on (i), input data 125 that is a candidate prescription is output.

[0120] As an example, the output unit 160 refers to the output result RS3 from the correction unit 158 ​​and calculates the maximum target achievement probability "P calc (i) and determine the target achievement probability "P calcOutput the input data 125 corresponding to "(i)" as a prescription candidate.

[0121] As another example, the output unit 160 refers to the output result RS3 by the correction unit 158, and determines the target achievement probability "P calc that falls within the top predetermined number, and identifies "(i)", and outputs the input data 125 corresponding to the target achievement probability "P calc (i)" as a prescription candidate.

[0122] The method of outputting the prescription candidate is arbitrary. As an example, the output unit 160 displays the formulation conditions defined in the input data 125 selected as the prescription candidate on the above-described display device 106. As another example, the output unit 160 stores the formulation conditions defined in the input data 125 selected as the prescription candidate in the above-described auxiliary storage device 120.

[0123] <F. Flowchart related to learning process> Next, referring to FIG. 11, the flow of the learning process by the information processing apparatus 100 will be described. FIG. 11 is a flowchart showing the flow of the learning process.

[0124] The control device 101 of the information processing apparatus 100 functions as the above-described learning unit 152 by executing the above-described learning program 128, and executes the learning process shown in FIG. 11. In other aspects, part or all of the learning process may be executed by circuit elements or other hardware.

[0125] In step S110, the control device 101 initializes the variable "j". As an example, the variable "j" is initialized to "1".

[0126] In step S112, the control device 101 acquires the j-th performance data 123 included in the learning data set 122.

[0127] In step S114, the control device 101 inputs the blending conditions as explanatory variables defined in the performance data 123 obtained in step S112 into the classification model 126. As a result, the classification model 126 outputs the probabilities of being classified into each evaluation label defined for the qualitative evaluation test.

[0128] In step S116, the control device 101 calculates the error between the classification probability as the correct value defined in each performance data 123 obtained in step S112 and the classification probability as the prediction result obtained in step S114, and updates the internal parameters of the classification model 126 so that the error becomes smaller than the current value. The parameters are updated, for example, by the error backpropagation method.

[0129] In step S120, the control device 101 determines whether to end the learning process. As an example, the control device 101 determines to end the learning process when the variable "j" is greater than a predetermined value. Alternatively, the control device 101 determines to end the learning process when all the performance data 123 included in the learning data set 122 have been used for learning. When the control device 101 determines to end the learning process (YES in step S120), it ends the process shown in FIG. 11. Otherwise (NO in step S120), the control device 101 switches the control to step S122.

[0130] In step S122, the control device 101 increments the variable "j". That is, the control device 101 increases the variable "j" by 1. Then, the control device 101 returns the control to step S112.

[0131] <G. Flowchart related to prediction processing> Next, referring to FIGS. 12 and 13, the flow of the prediction processing by the information processing device 100 will be described. FIGS. 12 and 13 are flowcharts showing the flow of the prediction processing.

[0132] The control device 101 of the information processing device 100 executes the prediction program 130 described above to perform the prediction process shown in Figures 12 and 13. In another aspect, part or all of the prediction process may be performed by circuit elements or other hardware.

[0133] 12 and 13, the prediction process is realized by executing steps S200A, S200B, and S260. Step S200A includes steps S210, S212, S214, S216, S220, and S222. Step S200B includes steps S230, S232, S234, S240, S242, S250, and S252.

[0134] In step S210, the control device 101 initializes a variable "i." As an example, the variable "i" is initialized to "1."

[0135] In step S212, the control device 101 functions as the achievement probability calculation unit 154 described above, and acquires the target condition 132 described above.

[0136] In step S214, the control device 101 functions as the achievement probability calculation unit 154 described above, and acquires the i-th input data 125 included in the prediction dataset 124. Next, the control device 101 inputs the acquired input data 125 to the classification model 126. As a result, the classification model 126 outputs the probability of being classified into each evaluation rank defined for the qualitative evaluation test.

[0137] In step S216, the control device 101 functions as the achievement probability calculation unit 154 described above, and calculates the target achievement probability "P pred (i) is calculated. The probability of achieving the target is P pred The calculation method for (i) is as described above, and therefore the explanation will not be repeated.

[0138] In step S220, the control device 101 calculates the target achievement probability "P pred As an example, when the variable "i" is greater than a predetermined value, the control device 101 determines whether or not to end the calculation process of the target achievement probability "P pred Alternatively, the control device 101 determines that the calculation process of "(i)" is to be terminated when all of the input data 125 included in the prediction data set 124 has been used in the calculation process. The control device 101 determines that the calculation process is to be terminated when all of the input data 125 included in the prediction data set 124 has been used in the calculation process. pred If it is determined that the calculation process of "(i)" is to be ended (YES in step S220), the process switches to step S230. Otherwise (NO in step S220), the control device 101 switches the control to step S222.

[0139] In step S222, the control device 101 increments the variable "i." That is, the control device 101 increases the variable "i" by 1. Thereafter, the control device 101 returns the control to step S212.

[0140] In step S230, the control device 101 initializes a variable "i." As an example, the variable "i" is initialized to "1."

[0141] In step S232, the control device 101 functions as the distance weight calculation unit 156 described above, and acquires the i-th input data 125 included in the prediction dataset 124. Next, the control device 101 calculates the distance between each piece of performance data 123 included in the training dataset 122 and the i-th input data 125. The method for calculating the distance is as described above, and therefore the description thereof will not be repeated.

[0142] In step S234, the control device 101 functions as the distance weight calculation unit 156 described above, and identifies the smallest distance "d0(i)" from among the distances calculated in step S232.

[0143] In step S240, the control device 101 determines whether or not to end the calculation process of the minimum distance "d0(i)". The conditions for ending this calculation process are the same as those in step S220. If the control device 101 determines that the calculation process of the minimum distance "d0(i)" should be ended (YES in step S240), the control device 101 switches the process to step S250. If not (NO in step S240), the control device 101 switches the control to step S242.

[0144] In step S242, the control device 101 increments the variable "i." That is, the control device 101 increases the variable "i" by 1. Thereafter, the control device 101 returns the control to step S232.

[0145] In step S250, the control device 101 functions as the distance weight calculation unit 156 described above, normalizes the minimum distance "d0(i)" so that it is greater than or equal to 0 and less than or equal to 1, and calculates the normalized distance "d(i)." The normalization process is as described above, and therefore will not be described again.

[0146] In step S252, the control device 101 functions as the correction unit 158 ​​described above, and corrects the target achievement probability "P" calculated in step S216 with a weight according to the normalized distance "d(i)" calculated in step S250. pred As a result, the control device 101 calculates the weighted target achievement probability "P calc (i)" is calculated. The calculation process is as described above, and therefore the description thereof will not be repeated.

[0147] In step S260, the control device 101 functions as the output unit 160 described above, and outputs the weighted goal achievement probability "P calc Based on "(i)", input data 125 that will be prescription candidates is output. The prescription candidate output process is as described above, and therefore the description thereof will not be repeated.

[0148] [Second embodiment] <H. Summary> Next, referring to FIG. 14, the prediction process of the goal achievement probability according to the second embodiment will be described. FIG. 14 is a diagram schematically showing the prediction process according to the second embodiment.

[0149] The information processing apparatus 100 according to the first embodiment predicted candidates for an optimal prescription based on the goal achievement probability predicted for a single qualitative evaluation test. In contrast, the information processing apparatus 100 according to the second embodiment predicts candidates for an optimal prescription based on each goal achievement probability predicted for different types of qualitative evaluation tests.

[0150] As an example, the information processing apparatus 100 according to the present embodiment uses a plurality of classification models 126, 226. The classification models 126, 226 are generated in advance by a learning process using the learning dataset 122A shown in FIG. 15. FIG. 15 is a diagram showing the learning dataset 122A according to the second embodiment.

[0151] The learning dataset 122A shown in FIG. 15 is different from the above-described learning dataset 122 (see FIG. 3) in that it includes the evaluation ranks given in each of a plurality of types of qualitative evaluation tests as explanatory variables.

[0152] More specifically, the learning dataset 122A includes a plurality of performance data 123A. The number of performance data 123A included in the learning dataset 122A is arbitrary. As an example, the number of performance data 123A is several to tens of thousands.

[0153] Each of the performance data 123A associates an explanatory variable with a target variable. The value of the explanatory variable defined in the performance data 123A and the value of the target variable defined in the performance data 123A are measured values.

[0154] The explanatory variable defined in the performance data 123A is, for example, the blending condition of the raw material. Since the description of the explanatory material is as described above, the description thereof will not be repeated.

[0155] The objective variables defined in the performance data 123A include an evaluation rank assigned in the qualitative evaluation test "A" and an evaluation rank assigned in the qualitative evaluation test "B." The evaluation test "A" and the evaluation test "B" are different types of tests. Examples of qualitative evaluation tests are as described above, and therefore will not be described again.

[0156] The classification model 126 is generated by machine learning the relationship between the blending conditions defined in the performance data 123A and the evaluation ranks assigned in the qualitative evaluation test "A." The information processing device 100 inputs input data 125 including the blending conditions of raw materials into the classification model 126. As a result, the classification model 126 outputs the probability that a material produced under the blending conditions defined in the input data 125 will be classified into each evaluation rank in the evaluation test "A."

[0157] Next, the information processing device 100 calculates the goal achievement probability based on each classification probability obtained from the classification model 126, and weights the goal achievement probability. The calculation process of the goal achievement probability and the weighting process are as described above, and therefore the description thereof will not be repeated.

[0158] On the other hand, classification model 226 is generated by machine learning the relationship between the blending conditions defined in performance data 123A and the evaluation ranks assigned in qualitative evaluation test "B." Information processing device 100 inputs input data 125 including the blending conditions of raw materials to classification model 226. As a result, classification model 226 outputs the probability that a material produced under the blending conditions defined in input data 125 will be classified into each evaluation rank in evaluation test "B."

[0159] Next, the information processing device 100 calculates the goal achievement probability based on each classification probability obtained from the classification model 226, and weights the goal achievement probability. The calculation process of the goal achievement probability and the weighting process are as described above, and therefore the description thereof will not be repeated.

[0160] Next, the information processing apparatus 100 integrates the weighted target achievement probability calculated from the output value of the classification model 126 and the weighted target achievement probability calculated from the output value of the classification model 226. As an example, the information processing apparatus 100 integrates the respective target achievement probabilities by multiplying the respective target achievement probabilities. Thereby, the information processing apparatus 100 can specify, as a candidate for a prescription, a formulation condition that satisfies both the target achievement probability that satisfies the target condition in the evaluation test "A" and the target achievement probability that satisfies the target condition in the evaluation test "B".

[0161] Note that, in the above description, an example in which the classification probability related to the evaluation test "A" is output from the classification model 126 and the classification probability related to the evaluation test "B" is output from the classification model 226 has been described. However, the classification probability related to the evaluation test "A" and the classification probability related to the evaluation test "B" may be output from the same classification model.

[0162] Also, in the above description, an example using two classification models 126 and 226 has been described. However, three or more classification models may be used. As an example, N (N is an integer greater than or equal to 3) classification models learned for N evaluation tests may be used. In this case, the information processing apparatus 100 calculates the weighted target achievement probability for each of the classification models, and integrates the respective target achievement probabilities by multiplying the respective target achievement probabilities.

[0163] <I. Functional Configuration> Next, referring to FIGS. 16 to 20, the functional configuration of the information processing apparatus 100 according to the second embodiment will be described. FIG. 16 is a diagram showing an example of the functional configuration of the information processing apparatus 100 according to the second embodiment.

[0164] As shown in FIG. 16, the information processing apparatus 100 includes, as functional components, a learning unit 152, an achievement probability calculation unit 154, a distance weight calculation unit 156, a correction unit 158, a learning unit 252, an achievement probability calculation unit 254, a correction unit 258, an integration unit 259, and an output unit 260.

[0165] 16 does not necessarily have to be implemented in the information processing device 100. A part of the functional configuration shown in Fig. 16 may be implemented in the information processing device 100, and the remaining functional configuration may be implemented in another computer such as a server.

[0166] The functions of learning unit 152, achievement probability calculation unit 154, distance weight calculation unit 156, and correction unit 158 ​​are as described above, and therefore their description will not be repeated. Below, the functional configurations of learning unit 252, achievement probability calculation unit 254, correction unit 258, integrating unit 259, and output unit 260 will be described in order.

[0167] (I1. Learning Section 252) First, the function of the learning unit 252 shown in Fig. 16 will be described with reference to Fig. 17. Fig. 17 is a diagram conceptually showing the learning process performed by the learning unit 252.

[0168] The learning unit 252 executes a learning process using the above-described learning dataset 122A to generate the classification model 226. The machine learning algorithm employed is not particularly limited, and various machine learning algorithms such as a neural network such as deep learning, a support vector machine, or a decision tree system may be employed. The learning process using a neural network will be described below.

[0169] The classification model 226 is composed of an input layer X, a hidden layer H, and an output layer Y.

[0170] The input layer X is configured to receive input of explanatory variables defined in the performance data 123 A. As an example, the input layer X includes a unit group x1 and a unit group x2.

[0171] The unit group x1 is configured to receive explanatory variables (e.g., raw material blending ratios) defined in the performance data 123A. The number of units constituting the unit group x1 is the same as the number of dimensions of the blending ratios. As an example, if the blending ratios are five-dimensional, the unit group x1 is composed of five units. Each unit constituting the unit group x1 outputs the input data to each unit in the first layer of the intermediate layer H.

[0172] The unit group x2 is configured to receive input of other explanatory variables (e.g., manufacturing conditions) defined in the classification model 226. The number of units constituting the unit group x2 is the same as the number of dimensions of the manufacturing conditions. As an example, if the manufacturing conditions are expressed in two dimensions, the unit group x2 is composed of two units. Each unit constituting the unit group x2 outputs the input data to each unit in the first layer of the intermediate layer H.

[0173] The middle layer H is composed of multiple layers. The number of layers in the middle layer H is arbitrary. In the example of FIG. 17, the middle layer H is composed of N layers (N is a natural number). Each layer in the middle layer H includes multiple units. In the example of FIG. 17, the first layer of the middle layer H is composed of unit h A1 ,h A2 The final layer of the intermediate layer H is made up of units h N1 ,h N2 It is composed of...

[0174] Each unit constituting each layer of the hidden layer H is connected to each unit in the previous layer and each unit in the next layer. Each unit in each layer receives each output value from each unit in the previous layer, multiplies each output value by a weight, accumulates the multiplication results, adds (or subtracts) a predetermined bias to (or from) the accumulation result, inputs the addition result (or subtraction result) to a predetermined function (for example, a sigmoid function), and outputs the output value of the function to each unit in the next layer.

[0175] The output layer Y outputs an evaluation result (i.e., classification probability) according to the explanatory variables input to the input layer X. As an example, the output layer Y is composed of units y1 to y3. Hereinafter, the units y1 to y3 will also be referred to as unit y.

[0176] Each unit y is connected to each unit h in the final layer of the hidden layer H. N1 ,h N2 Each of the units y receives an output value from each unit in the final layer of the hidden layer H, multiplies each output value by a weight, accumulates the results of these multiplications, adds (or subtracts) a predetermined bias to (or from) the accumulated result, inputs the result of the addition (or subtraction) to a predetermined function (for example, a sigmoid function), and outputs the output result of the function as an output value.

[0177] The number of units constituting the output layer Y is determined according to the number of objective variables defined in the performance data 123A. As described above, the objective variables are the evaluation ranks given in the qualitative evaluation test "B," and the number of units constituting the output layer Y is, for example, the same as the number of evaluation ranks.

[0178] Unit y1 is configured to output a classification probability belonging to evaluation rank "1." Unit y2 is configured to output a classification probability belonging to evaluation rank "2." Unit y3 is configured to output a classification probability belonging to evaluation rank "3."

[0179] Note that the output value of the classification model 226 does not need to be the probability itself, but may be a value correlated with the probability. As an example, the output value of the classification model 226 may be a score correlated with the classification probability. In this case, the information processing device 100 normalizes the scores so that the sum of the scores becomes 1, and calculates the normalized score as the classification probability.

[0180] Furthermore, in the example of FIG. 17, one classification model 226 is configured to output classification probabilities belonging to evaluation ranks "1" to "3", but one classification model 226 may output one probability. As an example, the first classification model 226 is configured to receive input of the blending conditions of the ingredients and output the probability of belonging to evaluation rank "1". The second classification model 226 is configured to receive input of the blending conditions of the ingredients and output the probability of belonging to evaluation rank "2". The third classification model 226 is configured to receive input of the blending conditions of the ingredients and output the probability of belonging to evaluation rank "3".

[0181] Next, the update process of the internal parameters of the classification model 226 by the learning unit 252 will be described.

[0182] The learning unit 252 acquires the first piece of performance data 123A from the training dataset 122A, and inputs the explanatory variables defined in the performance data 123A to the classification model 226. As a result, the classification model 226 outputs the probability of belonging to each evaluation rank. Next, the learning unit 252 compares the output prediction results "p1" to "p3" with the objective variables defined in the first piece of performance data 123A. The objective variables are expressed, for example, as probabilities "YA" to "YC."

[0183] As an example, when the evaluation rank associated with the performance data 123A is "1," the value of the objective variable is (YA, YB, YC) = (1, 0, 0). When the evaluation rank associated with the performance data 123A is "2," the value of the objective variable is (YA, YB, YC) = (0, 1, 0). When the evaluation rank associated with the performance data 123A is "3," the value of the objective variable is (YA, YB, YC) = (0, 0, 1).

[0184] The learning unit 252 calculates the error "Z" between the output results "p1" to "p3" of the classification model 226 and the objective variables "YA" to "YC". Next, the learning unit 252 updates various parameters (e.g., weights and biases) included in the classification model 226 so as to reduce the error "Z". The parameter update is realized, for example, by the backpropagation method.

[0185] The learning unit 252 repeatedly updates the internal parameters of the classification model 226 for each piece of performance data 123A included in the learning data set 122A. As a result, the classification model 226 begins to output accurate prediction results as the learning progresses.

[0186] The learning unit 252 does not need to use all of the performance data 123A included in the training data set 122A in the training process, and may use a portion of the performance data 123A included in the training data set 122A to generate the classification model 226. The remaining performance data 123A is used, for example, to evaluate the classification model 226.

[0187] (I2. Achievement Probability Calculation Unit 254) Next, the function of the achievement probability calculation unit 254 shown in Fig. 16 will be described with reference to Fig. 18. Fig. 18 is a diagram conceptually showing the prediction process by the achievement probability calculation unit 254.

[0188] The achievement probability calculation unit 254 receives input of desired target conditions 232 related to the physical properties of the material. The target conditions 232 are input, for example, to an input screen displayed on the above-mentioned display device 106. The evaluator inputs the target conditions 232 to the input screen using, for example, the above-mentioned input device 108. Note that the target conditions 232 are different from the above-mentioned target conditions 132.

[0189] The target condition 232 is a condition for specifying a range of the evaluation rank associated with the output layer of the classification model 226. As an example, the target condition 232 is defined by at least one of a lower limit value of the evaluation rank and an upper limit value of the evaluation rank.

[0190] Next, the achievement probability calculation unit 254 acquires the above-mentioned prediction dataset 124 (see FIG. 4 ) and sequentially inputs the input data 125 included in the prediction dataset 124 to the classification model 226. As a result, the classification model 226 outputs the probability of being classified into each evaluation rank defined for the qualitative evaluation test “B.”

[0191] When a numerical range equal to or less than a predetermined threshold is specified as the goal condition 232, the achievement probability calculation unit 254 sums up the classification probabilities for the evaluation ranks equal to or less than the threshold, and calculates the sum result as the goal achievement probability "P' pred The "i" is the number of the input data 125 defined in the prediction dataset 124. As an example, when the input target condition 232 is an evaluation rank of "2" or less, the achievement probability calculation unit 254 calculates the target achievement probability "P'(i)" as the sum of the classification probability "p1" for the evaluation rank "1" and the classification probability "p2" for the evaluation rank "2". pred (i)" is calculated as follows.

[0192] When a numerical range equal to or greater than a predetermined threshold is specified as the goal condition 232, the achievement probability calculation unit 254 sums up the classification probabilities for evaluation ranks equal to or greater than the threshold, and calculates the sum as the goal achievement probability "P' pred As an example, when the input target condition 232 is an evaluation rank of "2" or higher, the achievement probability calculation unit 254 calculates the target achievement probability "P'(i)" as the sum of the classification probability "p2" for the evaluation rank "2" and the classification probability "p3" for the evaluation rank "3". pred (i)" is calculated as follows.

[0193] When a numerical range equal to or greater than the first threshold and equal to or less than the second threshold is specified as the goal condition 232, the achievement probability calculation unit 254 sums up the classification probabilities for the evaluation ranks that are equal to or greater than the first threshold and equal to or less than the second threshold, and calculates the sum result as the goal achievement probability "P' pred (i)" is calculated as follows.

[0194] (I3. Correction unit 258) Next, the function of the correction unit 258 shown in Fig. 16 will be described with reference to Fig. 19. Fig. 19 is a diagram conceptually showing the weighting process performed by the correction unit 258.

[0195] 19 shows an output result RS1A from the achievement probability calculation section 254, an output result RS2 from the distance weight calculation section 156, and an output result RS3A from the correction section 258. In FIG.

[0196] The correction unit 258 performs the same weighting process as the correction unit 158 ​​described above. More specifically, the correction unit 258 adjusts the target achievement probability "P'" calculated by the achievement probability calculation unit 254 with a weight according to the distance "d(i)" calculated by the distance weight calculation unit 156 described above. pred As an example, the correction unit 258 weights the target achievement probability "P' pred (i)" and the distance "d(i)" specified in the output result RS2 to obtain "P' calc (i)" is calculated.

[0197] The output result RS3A from the correction unit 258 is output to the integration unit 259.

[0198] (I4. Integration Section 259) Next, the function of the integrating unit 259 shown in Fig. 16 will be described with reference to Fig. 20. Fig. 20 is a diagram conceptually showing the integrating process performed by the integrating unit 259.

[0199] The integration unit 259 integrates the output result RS3 by the correction unit 158 ​​and the output result RS3A by the correction unit 258. As an example, the integration unit 259 integrates the weighted goal achievement probability "P calc (i)" and the weighted target achievement probability "P'" specified in the output result RS3A. calc (i)" to calculate the integrated result RS4.

[0200] (I5. Output section 260) Continuing to refer to FIG. 20, the function of the output unit 260 shown in FIG. 16 will be described.

[0201] Based on the integration result RS4 by the integration unit 259, the output unit 260 outputs the input data 125 that is a candidate for the prescription.

[0202] As an example, the output unit 260 refers to the integration result RS4 by the integration unit 259, and determines the maximum target achievement probability "P calc (i)", and outputs the input data 125 corresponding to the target achievement probability "P calc (i)" as a candidate for the prescription.

[0203] As another example, the output unit 260 refers to the integration result RS4 by the integration unit 259, and determines the target achievement probabilities "P calc (i)" that fall within the top predetermined number, and outputs the input data 125 corresponding to the target achievement probabilities "P calc (i)" as candidates for the prescription.

[0204] The method of outputting the prescription candidate is arbitrary. As an example, the output unit 260 displays the formulation conditions defined in the input data 125 selected as the prescription candidate on the above-described display device 106. As another example, the output unit 260 stores the formulation conditions defined in the input data 125 selected as the prescription candidate in the above-described auxiliary storage device 120.

[0205] <J. Others> Note that the above-described integration unit 259 integrated each target achievement probability by multiplying the target achievement probabilities for each physical property equally. However, depending on the physical property to be evaluated, it may be more important than other physical properties. Therefore, the integration unit 259 may integrate each target achievement probability after weighting the target achievement probabilities for each physical property. ​​​​​​​​​calc_1 (i))·w1+log(P calc_2 (i))·w2+···+log(P calc_n (i))·w n ···(7) The log(P calc_all (i))" indicates the logarithm of the integrated probability of achieving the target for all properties after weighting. calc_n (i))" indicates the logarithm of the target achievement probability for the nth physical property. n " indicates the weight set for the nth physical property. "n" indicates a positive integer. "w n " may be arbitrarily set by the evaluator or may be set in advance. Typically, "w n The sum of " is 1.

[0208] The embodiments disclosed herein should be considered to be illustrative in all respects and not restrictive. The scope of the present invention is defined by the claims, not by the above description, and is intended to include all modifications within the meaning and scope of the claims. [Explanation of symbols]

[0209] 100 Information processing device, 101 Control device, 102 ROM, 103 RAM, 104 Communication interface, 105 Display interface, 106 Display device, 107 Input interface, 108 Input device, 115 Bus, 120 Auxiliary storage device, 122 Learning dataset, 122A Learning dataset, 123 Actual data, 123A Actual data, 124 Prediction dataset, 125 Input data, 126 Classification model, 128 Learning program, 130 Prediction program, 132 Target condition, 152 Learning unit, 154 Achievement probability calculation unit, 156 Distance weight calculation unit, 158 Correction unit, 160 Output unit, 226 Classification model, 232 Target condition, 252 Learning unit, 254 Achievement probability calculation unit, 258 Correction unit, 259 Integration unit, 260 Output unit, RS1 Output result, RS1A Output results, RS2 output results, RS3 output results, RS3A output results, RS4 integrated results.

Claims

1. A method performed by an information processing device, comprising: A method for producing a material using a first evaluation test comprising: inputting input data including raw material blending conditions into a first classification model; and acquiring a probability that a material produced under the blending conditions will be classified into each of a plurality of evaluation results defined for a predetermined first evaluation test; the first classification model is generated by a learning process using a plurality of first performance data, each of the plurality of first performance data associates blending conditions of raw materials with evaluation results of the first evaluation test conducted on the material produced under the blending conditions; a step of predicting a first target achievement probability that a material produced under the blending conditions defined in the input data will satisfy a desired condition, based on the plurality of probabilities acquired in the acquiring step; For each of the plurality of first performance data, a step of calculating a similarity between a compounding condition defined in the first performance data and a compounding condition defined in the input data; weighting the first goal achievement probability with a first weight according to the plurality of similarities calculated in the calculating step.

2. The method further comprises: weighting the first goal achievement probability weighted by the first weight with a second weight; 2. The method of claim 1, further comprising: outputting a result of adding the first goal achievement probability weighted with the first weight and the second weight and the first goal achievement probability not weighted with the first weight and the second weight.

3. The method further includes weighting the first goal achievement probability, which has not been weighted with the first weight, with a third weight; The method of claim 2 , wherein the sum is a sum of the first goal achievement probability weighted by the first weight and the second weight and the first goal achievement probability weighted by the third weight.

4. The method according to claim 3 , further comprising the step of decreasing the second weight as the number of the plurality of performance data increases.

5. The method according to claim 4 , further comprising the step of increasing the third weight as the number of the plurality of performance data increases.

6. The method according to any one of claims 3 to 5, wherein the sum of the second weight and the third weight is constant.

7. In the obtaining step, a plurality of pieces of input data are input to the first classification model; The method according to any one of claims 2 to 5, further comprising the step of outputting, from among the plurality of input data, the input data that produces the largest addition result.

8. The method according to any one of claims 2 to 5, wherein the first weight is calculated based on the similarity of the performance data that is most similar to the input data among the plurality of performance data.

9. The method according to any one of claims 1 to 5, wherein the blending conditions include at least one of a blending ratio of raw materials and a manufacturing condition when blending the raw materials.

10. The method of any one of claims 1 to 5, wherein the material comprises a sticky adhesive material.

11. The method further includes a step of inputting the input data including blending conditions of raw materials into a second classification model, and obtaining a probability that a material produced under the blending conditions will be classified into each of a plurality of evaluation results defined for a predetermined second evaluation test; the second classification model is generated by a learning process using a plurality of second performance data, each of the plurality of second performance data associates blending conditions of raw materials with evaluation results of the second evaluation test conducted on the material produced under the blending conditions; The method further comprises: predicting a second target achievement probability that a material produced under formulation conditions defined in the input data will satisfy a desired condition based on the plurality of probabilities obtained from the second classification model; For each of the plurality of second performance data, calculating a similarity between the compounding conditions defined in the second performance data and the compounding conditions defined in the input data; weighting the second goal achievement probability with a first weight according to the plurality of similarities calculated for the input data; The method according to any one of claims 1 to 5, further comprising the step of aggregating the first goal achievement probability and the second goal achievement probability.

12. An information processing device, A control unit is provided, the control unit inputs input data including blending conditions of raw materials into a first classification model, and executes a process of acquiring a probability that a material produced under the blending conditions will be classified into each of a plurality of evaluation results defined for a predetermined first evaluation test; the first classification model is generated by a learning process using a plurality of first performance data, each of the plurality of first performance data associates blending conditions of raw materials with evaluation results of the first evaluation test conducted on the material produced under the blending conditions; The control unit further a process of predicting a first target achievement probability that a material produced under the blending conditions defined in the input data will satisfy a desired condition, based on the plurality of probabilities acquired in the acquiring process; For each of the plurality of first performance data, a process of calculating a similarity between a compounding condition defined in the first performance data and a compounding condition defined in the input data; and weighting the first goal achievement probability with a first weight according to the plurality of similarities calculated in the calculating process.

13. A program executed by an information processing device, The program is configured to: inputting input data including blending conditions of raw materials into a first classification model, and executing a process of acquiring the probability that a material produced under the blending conditions will be classified into each of a plurality of evaluation results specified for a predetermined first evaluation test; the first classification model is generated by a learning process using a plurality of first performance data, each of the plurality of first performance data associates blending conditions of raw materials with evaluation results of the first evaluation test conducted on the material produced under the blending conditions; The program further includes: a process of predicting a first target achievement probability that a material produced under the blending conditions defined in the input data will satisfy a desired condition, based on the plurality of probabilities acquired in the acquiring process; For each of the plurality of first performance data, a process of calculating a similarity between a compounding condition defined in the first performance data and a compounding condition defined in the input data; and weighting the first goal achievement probability with a first weight according to the plurality of similarities calculated in the calculation process.

Images