Model generation system and model generation method

Abstract

The present invention comprises: a source model database 12 that stores a source model; and a model generation unit 11 that generates a target model by using the source model searched from the source model database. The model generation unit comprises: a database search unit that searches for a first source model 31 which includes, as an output thereof, an output of the target model, and a second source model 32 which includes, as an input thereof, an input of the target model; and a connection determination unit that, when it is possible to associate the input of the first source model with the output of the second source model, connects the input of the first source model with the output of the second source model.

Description

[Technical field]

[0001] The present invention relates to a model generation system and a model generation method. [Background technology]

[0002] With the recent advances in information processing technology, the use of machine learning technology is increasing in various fields. Machine learning is a method or technology that constructs a learning model (also simply called a "model") from a large amount of data or experience with certain rules, and uses the learning model to execute some task.

[0003] In general, in machine learning, the more data used for learning (called "learning data" or "training data"), the more accurate a learned model can be realized. Conversely, when there is little data, it is difficult to build a learned model with sufficient accuracy. Therefore, when building a learning model, if a lot of learning data is obtained, a learned model with sufficient accuracy can be obtained, but if there is little learning data, only a learned model with insufficient accuracy can be obtained.

[0004] In the field of machine learning technology, there is a technique called transfer learning for obtaining a highly accurate model with a small amount of data. Transfer learning is a general term for techniques that attempt to achieve sufficient accuracy even when there is a small amount of training data by utilizing (transferring) a trained model created using other training data. In transfer learning, attempts are made to obtain a highly accurate model by reusing an existing trained model as it is, retraining using a trained model as an initial value (such as "fine tuning"), or incorporating a trained model as a partial model (source model) as part of a target model to construct a new model. For example, Patent Document 1 discloses a method for selecting a trained model to be used in transfer learning. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2021-182329 Summary of the Invention [Problem to be solved by the invention]

[0006] Transfer learning techniques are expected to reduce the costs of preparing datasets and the computational costs involved in building large-scale, complex models by using previously created trained models.

[0007] On the other hand, various trained models have been constructed with the advancement of manufacturing technology and the commoditization of machine learning technology. Therefore, the present inventors have considered combining multiple trained models, such as black box models such as neural networks (NNs), and consolidating them into one model to generate a desired trained model. This makes it possible to obtain a model with high prediction accuracy while ensuring development efficiency (calculation amount and time required for learning) in model development. Furthermore, by training (transfer learning) a learning model in which trained models are combined in this way, it is possible to further improve prediction accuracy.

[0008] The object of the present invention is to make it possible to generate a learning model by efficiently utilizing similar past trained models, partial models, known physical equations, etc., without overlooking them, particularly in the development of a new model using a transfer learning method. [Means for solving the problem]

[0009] A model generation system according to one embodiment of the present invention is a model generation system for generating a target model, the model generation system including: a source model database for storing a source model; and a model generation unit for generating a target model by using a source model searched from the source model database; the model generation unit includes a database search unit that searches for a first source model including an output of the target model as its output and a second source model including an input of the target model as its input, and a combination determination unit that combines the input of the first source model with the output of the second source model when the input of the first source model can be associated with the output of the second source model; Source models are stored in the database as source models, including trained machine learning models. The combination determination unit determines that association is possible only when an item name of an input of the first source model matches an item name of an output of the second source model, when an item name of an input of the first source model and an item name of an output of the second source model are synonymous, when there is a combination history between an item name of an input of the first source model and an item name of an output of the second source model, or when it is determined that dimensional information of a unit of an input of the first source model and dimensional information of a unit of an output of the second source model are similar. . Effect of the Invention

[0010] The learning cost (data collection cost, calculation cost) required for model generation is reduced. Other objects and novel features will become apparent from the description of this specification and the accompanying drawings. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram for explaining a basic configuration of a model generation system. [Diagram 2] 1 is an example of a process of a processing device predicted by a machine learning model. [Diagram 3] 2 is an example of a model generated in the first embodiment. [Figure 4] 2 is an example of a hardware configuration of an information processing device. [Diagram 5] 2 illustrates an example of programs and data stored in a storage device in the first embodiment. [Figure 6] 4 is a flowchart showing a process of generating a model by the model generation system in the first embodiment. [Figure 7] 13 is an example of a model generated in the second embodiment. [Figure 8] 13 is an example of programs and data stored in a storage device in the second embodiment. [Figure 9] 11 is a flowchart showing a process of generating a model by the model generation system in the second embodiment. [Figure 10]13 is an example of a model generated in the third embodiment. [Figure 11] 13 is an example of programs and data stored in a storage device in the third embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. However, the present invention is not to be interpreted as being limited to the description of the embodiment shown below. It is easily understood by those skilled in the art that the specific configuration can be changed without departing from the idea or intent of the present invention. In addition, the position, size, shape, etc. of each configuration shown in the drawings in this specification may not represent the actual position, size, shape, etc. in order to facilitate understanding of the invention. Therefore, the present invention is not limited to the position, size, shape, etc. disclosed in the drawings, etc.

[0013] The basic configuration of a model generation system 10 of this embodiment will be described with reference to Fig. 1. The model generation system 10 includes a model generation unit 11 and a source model database 12.

[0014] The source model database 12 stores source models to be used for constructing a target model. The source models stored in the database 12 include trained machine learning models (hereinafter referred to as trained models) and equations and inequalities. The format of the machine learning model is not limited, and machine learning models such as neural networks (NN), gradient boosting trees, linear regression, and kernel ridge methods are included. The equations and inequalities may be equations and inequalities that explain the phenomenon, and include various equations such as Newton's equation of motion (F=ma) and Langmuir's adsorption isotherm (K=Nθ / N(1-θ)p), as well as various inequalities such as Chebyshev's inequality and Clausius' inequality. The equations and inequalities may also be user-defined equations and inequalities. However, the equations stored in the source model database 12 are such that one parameter of the equation is treated as output Y (objective variable) and the remaining parameters are all treated as input X (explanatory variables or constants) so that the value of the objective variable is uniquely determined. The user can arbitrarily determine which parameter is to be output Y. Furthermore, for an inequality stored in the source model database 12, all parameters of the inequality can be treated as input X, and information (Boolean value) of 1 or 0 indicating whether the inequality is satisfied or not can be treated as output Y. If an inequality has an equality condition, it can be treated equivalently to an equation, that is, one parameter can be treated as output Y.

[0015] The model generation unit 11 combines trained models and equations / inequalities stored in the source model database 12. A source model has one or more inputs X and outputs Y, and the model generation unit 11 checks the item names of the inputs X and outputs Y of different source models, and combines the inputs X and outputs Y when the item names can be associated with each other.

[0016] FIG. 1 shows an example in which trained models 15 to 17 are stored in the source model database 12. The model generation unit 11 generates, for example, the output y 1,1 is the input x of trained model 16 2,1 Since it matches, the output y1,1 and type x 2,1 In addition, the model generation unit 11 combines, for example, the output y 1,1 and the input x of trained model 17 3,1 Since there is no correspondence between y and y, the output 1,1 and type x 3,1 The cases where the model generating unit 11 can associate item names with each other include cases where the item names match and cases where some correspondence is recognized between the item names. A specific method for determining whether or not a combination is possible will be described later.

[0017] As a rule for combination by the model generation unit 11, one output Y of a source model can be combined with inputs X of multiple other source models, whereas one input X of a source model can only be combined with one output Y of other source models. Also, basically, inputs and outputs of multiple source models cannot be combined in a loop. EXAMPLES

[0018] In Example 1, an example is shown in which a machine learning model for predicting the processing results of a processing device is constructed using a first trained model trained with data acquired by actually operating the processing device, and a second trained model trained with the results obtained by simulating physical phenomena occurring in the processing device using computer software, using the model generation system of this example.

[0019] As shown in Fig. 2, raw material 21 is fed into processing device 22 to obtain product 23. Here, the state of product 23 depends on what processing conditions are set for processing device 22 by control computer 24. Therefore, processing conditions that can be set for processing device 22 are comprehensively assigned, raw material 21 is actually processed in processing device 22, a large amount of learning data is obtained regarding what product 23 was obtained under each processing condition, and a machine learning model (e.g., a neural network model) is trained to obtain a learned model that predicts the product state from the processing conditions with a certain degree of accuracy.

[0020] However, since it usually takes a long time for the processing device 22 to process the raw materials 21, it takes a long time to obtain a large amount of new learning data. Therefore, in the model generation system of this embodiment, a method is described for obtaining a machine learning model that can predict the product state without obtaining new learning data by utilizing knowledge and simulation technology obtained in the past.

[0021] For example, as shown in FIG. 3, it is assumed that the user intends to construct a target model 33 in which five control parameters, namely, "gas pressure", "coil current", "power", "element ratio", and "voltage", which are independent processing conditions of the processing device 22, are input as X, and the "product state" of the product 23 is output as Y. It is also assumed that a first source model 31 and a second source model 32 are stored in the source model database 12. The first source model 31 is a learned model that has been trained using the product state of the product 23 obtained by making the processing device 22 process the raw material 21 by comprehensively assigning three control parameters, "power", "element ratio", and "voltage", in addition to the "ion flow rate" in the processing device 22, as learning data. The ion flow rate in the device can be measured by installing a measuring device in the processing device 22. It is assumed that the first source model 31 has been constructed through past experiments and the like, in which four parameters, "ion flow rate", "power", "element ratio", and "voltage", are input as X, and the "product state" is output as Y, and stored in the source model database 12. The second source model 32 is constructed using a physical simulation technique for the physical phenomena occurring in the processing chamber of the processing device 22, and is stored in the source model database 12. The second source model 32 is a trained model in which four control parameters, "gas pressure", "coil current", "power", and "element ratio", are input X, and "ion flow rate" is output Y. In general, the calculation time required for the simulation is very short compared with the time required to actually process the raw material 21 in the processing device 22, and the cost associated with acquiring the training data can be reduced.

[0022] In this way, if physical simulation is possible, a large amount of training data can be prepared in a short time, so that it is possible to generate a highly accurate trained model using machine learning such as deep learning. Since it is trained with a large amount of training data, such a trained model is expected to predict the output Y ("ion flow rate" in the case of the second source model 32) with high accuracy.

[0023] The model generation system 10 obtains a desired target model 33 by combining a first source model 31 and a second source model 32 .

[0024] The model generation system 10 is realized by an information processing device 40 including a processor (CPU) 41, a memory 42, a storage device 43, an input device 44, an output device 45, a communication device 46, and a bus 47 as main components as shown in FIG. 4. The processor 41 functions as a functional unit (functional block) that provides a predetermined function by executing processing according to a program loaded in the memory 42. The storage device 43 stores data used in the functional unit in addition to programs that cause the functional unit to function. For the storage device 43, a non-volatile storage medium such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive) is used. The input device 44 is a keyboard, a pointing device, etc., and the output device 45 is a display, etc. The communication device 46 enables communication with other information processing devices via a network. These are connected to each other so that they can communicate with each other via the bus 47.

[0025] The model generation system 10 does not have to be realized by one information processing device, and may be realized by multiple information processing devices. Also, some or all of the functions of the model generation system 10 may be realized as an application on a cloud.

[0026] 5 shows the programs and data stored in the storage device 43. A model generation program 51 is loaded into the memory 42 and executed by the processor 41, causing the processor 41 to function as the model generation unit 11. The model generation program 51 includes, as subprograms, a database (DB) search program 52, a bond judgment program 53, and a model determination program 54. These subprograms are also loaded into the memory 42 and executed by the processor 41, causing the processor 41 to function as the DB search unit, the bond judgment unit, and the model determination unit. The source model database 12 used by the model generation system is also stored in the storage device 43.

[0027] A flowchart showing the process of generating the target model 33 using the model generation system 10 is shown in Fig. 6. First, the user sets the names of the input items and the output items of the target model he or she wishes to create (S01). In the example of Fig. 3, the names of the five control parameters (such as "gas pressure") that will be the input X of the model 33 and the parameter ("product state") that will be the output X are set.

[0028] Next, the DB search unit searches the source model database 12 for a source model having an input item name equal to the set input item name and a source model having an output item name equal to the set output item name (S02). If there are multiple candidates, the user may be presented with options to select from, or the system may select a model that is most recently updated. In the example of FIG. 3, it is assumed that a second source model 32 having an input item name such as "gas pressure" and an output item name such as "ion flow rate" and a first source model 31 having an input item name such as "ion flow rate" and "electric power" and an output item name such as "product state" have been searched for.

[0029] Next, the join determination unit joins the input and output names of the source models together (S03). Here, as a simple example, an example is shown in which input and output names are joined when they match. In the example of Fig. 3, the output item name "ion flow rate" of the second source model 32 matches the input item name "ion flow rate" of the first source model 31, so these are joined.

[0030] Next, the model determination unit displays the combined source models on the output device 45 (S04). At this time, for example, the model combination diagram 35 including the model combination information as shown in the dashed line frame of FIG. 3 is displayed on the output device 45. In the model combination diagram 35, an input node indicating an input X is displayed on the left side of a box indicating a source model, and an output node indicating an output Y is displayed on the right side. Furthermore, an input node (processing condition) indicating an input X of the target model is displayed on the further left side of the source model, and the combination between the input node of the target model and the corresponding input node of the source model, and the combination between the corresponding nodes of the source model are displayed by edges. In this way, by displaying the GUI having a layout in which the input node is relatively located on the left side and the output node is relatively located on the right side on the display screen, it becomes easy to check whether there is an inappropriate combination (for example, a loop connection) between the source models.

[0031] The user checks the model binding diagram 35 displayed on the GUI screen (S05), and if correction is required, manually corrects the edges in the model binding diagram 35 on the GUI screen to correct the connections between the input nodes of the target model and the source model, or between source models (S06). Cases where correction is required include, for example, cases where the connected nodes are judged to be inappropriate based on the user's domain knowledge. After that, the completed target model 33 is saved in the source model database 12 (S07).

[0032] The model 33 created in this way is created using only the trained model and does not necessarily require additional learning. However, if even a small amount of training data (here, a data set of processing conditions for the five control parameters of the target model 33 and the product state under the processing conditions) is obtained, it is recommended to perform additional training (referred to as additional learning) using the training data. Using the weights of the original trained model as initial values ​​and updating the weights by additional learning using a small amount of training data is called fine tuning. It is generally known that performing appropriate fine tuning increases the possibility of a more accurate training model. When performing fine tuning, the user may appropriately set hyperparameters such as the learning rate of each model. EXAMPLES

[0033] In the second embodiment, not only trained models, equations, and inequalities but also untrained machine learning models are utilized as source models. The model generation system 10 of the second embodiment is also realized by an information processing device 40 as shown in FIG. 4. FIG. 8 shows programs and data stored in a storage device 43. In addition to the programs and data stored in the first embodiment (FIG. 5), a machine learning program 81 for performing machine learning is stored. The machine learning program 81 is loaded into the memory 42 and executed by the processor 41 to cause the processor 41 to function as a machine learning unit. The machine learning program 81 includes a model setting program 82 and a learning (training) program 83 as subprograms. These subprograms are also loaded into the memory 42 and executed by the processor 41 to cause the processor 41 to function as a model setting unit and a learning unit.

[0034] 7, in order to further improve the accuracy of the model 33 created in the first embodiment, a target model 70 is created by increasing the input X (control parameters), and an untrained machine learning model is used as the source model. Specifically, the input X of the target model 70 is obtained by adding two control parameters, "frequency" and "duty ratio", to the input X of the model 33.

[0035] A flowchart showing the process of generating a target model 70 using the model generation system 10 is shown in FIG. 9. The same processes as those in FIG. 6 are assigned the same reference numerals, and duplicated explanations are omitted, with differences being mainly explained. First, the user sets the names of input items and output items of the target model to be created (S01). In the example of FIG. 7, the names of seven control parameters (such as "gas pressure") that will be the input X of the target model 70 and the parameter ("product state") that will be the output X are set.

[0036] Next, the user specifies the learning data to be used for model learning (training) (S11). Since this example is a regression problem, the seven control parameters that are the input X of the target model 70 are comprehensively assigned, and the product state of the product 23 obtained by processing the raw material 21 with the processing device 22 is used as the learning data. The learning data, which is a combination of the seven control parameters and the product state, is provided in the form of, for example, a csv file. After that, steps S02 and S13 are performed in parallel.

[0037] In step S02, the DB search unit searches the source model database 12, and it is assumed that the model 33 is found. In step S03, the combination determination unit combines the input and output names of the source model. In this example, one model is found from the source model database 12, but if two or more models are found, the same processing as in the first embodiment is performed.

[0038] Next, the model determination unit displays the combined source models on the output device 45 (S04). In the second embodiment as well, a model combination diagram 75 including model combination information as shown in the dashed-dotted frame in Fig. 7 is displayed on the output device 45. However, at this point, only the searched source model 33 is displayed on the output device 45, and the "Frequency" node and the "Duty ratio" node are not connected anywhere.

[0039] The user checks the model binding diagram 75 displayed on the GUI screen (S05), and since there are unconnected processing conditions, manually corrects the model binding diagram 75 on the GUI screen (S06). Here, two unlearned models 71, 72 are added, and the source models are bound to each other in accordance with the desired target model 70. Note that the method of adding and binding unlearned models is not limited to the example in FIG. 7.

[0040] Next, the machine learning unit uses the learning data specified in step S11 to execute learning of a model (called a "combined model") that is a combination of the source model 33 and the unlearned models 71, 72 (S12).

[0041] The model setting unit allows the user to define inputs and outputs and set hyperparameters for the unlearned models 71, 72 added on the GUI screen. For example, when a neural network (NN) model is used as the unlearned model, the user can directly set detailed settings for various hyperparameters including the number of layers and the number of nodes, or the hyperparameters can be automatically determined by a program that performs Bayesian optimization. When performing Bayesian optimization, detailed settings for the range of optimization are also possible.

[0042] The learning unit trains the combined model using the training data. In training the combined model, for example, if the accuracy of the source model 33 is known to be relatively high or the amount of training data specified in step S11 is small, it is recommended to set the learning rate of the source model 33 to 0 and not update the weighting inside the source model 33 in the training in step S12. In such a case, the user turns on the lock icon 73 displayed in the upper right corner of the corresponding source model on the GUI screen. In training the combined model, if the lock icon 73 of the source model 33 is turned on, the learning unit does not update the weighting. This state is called the "source fixed mode."

[0043] On the other hand, when the lock icon 73 is turned off, the learning unit performs updates during training of the combined model, including the weighting of the source model 33. This state is called the "fine tuning mode", and for example, the additional learning described in the first embodiment is performed in the fine tuning mode.

[0044] Meanwhile, the machine learning unit constructs a model (called a "standard model") with the desired seven control parameters as input X and the "product state" as output Y (S13). The model setting unit defines the input and output of the standard model and allows the user to set hyperparameters for the standard model, and the learning unit trains the standard model using training data.

[0045] Note that learning the model using the learning data (steps S12 and S13) may take several hours to several days depending on the amount of data and the specifications of the information processing device (computer).

[0046] When the training of the standard model and the combined model is completed, the model determination unit compares the cross-validation (CV) results of these two models that the training unit performed during model training using a portion of the training data (S14). The model determination unit determines which model is capable of more accurate prediction based on the CV value, selects the model determined to have the more accurate prediction as the target model 70, and stores it in the source model database 12 (S07). EXAMPLES

[0047] The third embodiment is an example of constructing a model including even more processing conditions as input X, and an example of utilizing a physical equation stored in the source model database 12 will be described. Also, a method of combining nodes whose item names do not completely match will be described. The model generation system 10 of the third embodiment is also realized by an information processing device 40 as shown in FIG. 4. FIG. 11 shows programs and data stored in the storage device 43. In addition to the programs and data stored in the second embodiment (FIG. 8), a synonym dictionary 112 and a node combination history 113 are stored. Also, the combination judgment program 53 includes a text mining program 111 as a subprogram. The text mining program 111 is also loaded into the memory 42 and executed by the processor 41, causing the processor 41 to function as a text mining unit.

[0048] 10 is an example of creating a model with more inputs X (control parameters) than the model created in Example 2 in order to further improve accuracy, and an unlearned machine learning model and a physical equation are used as source models. Specifically, the input X of the model created in Example 3 is obtained by adding two control parameters, "temperature" and "processing time", to the input X of model 70.

[0049] These two control parameters are added based on the user's domain knowledge. For example, it is assumed that the longer the processing time, the greater the effect on the product state that is proportional to the time. Furthermore, if the reaction rate of the chemical reaction assumed to occur during processing is known, it can be estimated that the product of the reaction rate and the processing time has a large effect on the product state.

[0050] The process of generating a model using the model generation system 10 is the same as the flowchart shown in FIG. 9. The following mainly describes the features. First, the user sets the input item names and output item names of the target model to be created (S01). In the example of FIG. 10, the names of nine control parameters (such as "gas pressure") that will be the input X of the desired target model and the parameter ("product state") that will be the output X are set. In the following step S02, the DB search unit searches the source model database 12 to search for the first source model 31 and the second source model 32. In step S03, the combination determination unit combines the first source model 31 and the second source model 32. In step S04, the model determination unit displays the combined source model on the output device 45. In the third embodiment as well, a model combination diagram 105 including model combination information as shown in the dashed-dotted frame in FIG. 10 is displayed on the output device 45. However, at this point, only the searched source models 31 and 32 are displayed on the output device 45, and the “Temperature” node, the “Processing Time” node, the “Frequency” node, and the “Duty Ratio” node are not connected anywhere.

[0051] The user checks the model binding diagram 105 displayed on the GUI screen (S05), and since there are unconnected processing conditions, manually corrects the model binding diagram 105 on the GUI screen (S06). First, as in the second embodiment, two unlearned models 103, 104 are added, and the source models are bound to the desired target model. At this stage, the "Temperature" node and the "Processing Time" node remain unconnected.

[0052] As described above, when constructing a model to predict the product state, the user wanted to use the concepts of "processing time", which is one of the processing conditions, and "reaction rate" of the chemical reaction during processing, based on domain knowledge about the processing device 22. Therefore, two parameters, "reaction rate" and "processing time", are manually fixed in the input items of the unlearned model 104 in the final stage in Figure 10. "Processing time" is one of the parameters that is the input X of the target model, but "reaction rate" is an intermediate node that is neither an input node nor an output node of the target model.

[0053] At this point, since the "temperature" node and the "reaction rate" node remain as unconnected nodes, the user searches for a source model having these item names in the source model database 12. As a result, the Arrhenius equation (reaction rate k=Aexp(-E / RT)), which is a physical equation, is displayed on the GUI screen as a source model candidate. The Arrhenius equation is a general relational equation (formula) that expresses the correlation between the parameters of "reaction rate", "temperature", "concentration", and "activation energy (activation E)". When using an equation as a source model, it is necessary to assign one of the parameters to the output Y and assign the other parameters as the input X. Therefore, the source model 101 based on the Arrhenius equation is a source model in which "reaction rate" is the output Y and "temperature", "concentration", and "activation E" are the input X. As a result, the connection determination unit connects the "temperature" node of the input X of the source model 101 to the "temperature" node, which is a processing condition, and connects the "reaction rate" node, which is the output Y of the source model 101, to the input node to which "reaction rate" of the unlearned model 104 is assigned. As a result, the "concentration" node and the "activation E" node of the input X of the source model 101 remain unconnected.

[0054] Since all input nodes of the source model must be connected to some node, the connection judgment unit starts the text mining program 111. The text mining unit searches the synonym dictionary 112 or the node connection history 113, and automatically executes a search for "concentration" and "activation E". For example, it is assumed that the node connection history 113 has a history of manually connecting the "concentration" node and the "pressure" node based on the domain knowledge that "the higher the pressure, the higher the collision frequency of gas molecules, and the local molecular concentration related to the chemical reaction is substantially increased". The connection judgment unit connects the "concentration" node and the "gas pressure" node based on the above information searched by the text mining unit. On the other hand, the "activation E" node of the source model 101 was not automatically connected by the connection judgment unit, so a constant node 102 is set and a constant is input from the constant node 102 to the source model 101, thereby completing the connection correction between the source models.

[0055] Here, an example was shown in which the synonym dictionary 112 and the node link history 113 were searched for the link destination, but a method of inferring the link destination using the dimensional information of the unit assigned to each node is also useful. For example, for pressure (Pa), the SI unit system is used to estimate m -1 kg s -2 It is possible to evaluate the similarity or dissimilarity between units by utilizing the fact that it can be expressed as follows. Furthermore, an item correspondence table may be provided in which item names that the user is allowed to combine in advance with each other may be stored.

[0056] It is also possible to search the source model database 12 not only for the search of the connection destination but also for similarity in node names and connection history in this way (step S02). In this case, since a large number of source models will be searched, it is desirable to present them to the user in order of priority based on, for example, the number of corresponding variables.

[0057] Above, several embodiments of the model generation system have been described. The components disclosed in the above embodiments can be appropriately combined as long as there is no contradiction in the combination. Furthermore, some components can be deleted from all the components shown in the above embodiments. [Explanation of symbols]

[0058] 10: model generation system, 11: model generation unit, 12: source model database, 15 to 17: trained model, 21: raw material, 22: processing device, 23: product, 24: control computer, 31: first source model, 32: second source model, 33: model, 35: model combination diagram, 40: information processing device, 41: processor (CPU), 42: memory, 43: storage device, 44: input device, 45: output device, 46: communication device, 47: bus, 51: model generation program ram, 52: database search program, 53: binding judgment program, 54: model determination program, 70: model, 71, 72: unlearned model, 73: lock icon, 75: model binding diagram, 81: machine learning program, 82: model setting program, 83: learning program, 101: source model, 102: constant node, 103, 104: unlearned model, 105: model binding diagram, 111: text mining program, 112: synonym dictionary, 113: node binding history.

Claims

1. A model generation system for generating a target model, comprising: a source model database for storing source models; a model generation unit that generates the target model by using a source model retrieved from the source model database, The model generation unit a database search unit for searching a first source model that has as its output the output of the target model, and a second source model that has as its input the input of the target model; a combination determination unit that combines an input of the first source model with an output of the second source model when the input of the first source model can be associated with an output of the second source model, The source model stored in the source model database includes a trained machine learning model, The coupling determination unit determines that a correspondence is possible when an item name of an input of the first source model matches an item name of an output of the second source model, when an item name of an input of the first source model and an item name of an output of the second source model are synonyms, when there is a coupling history between an item name of an input of the first source model and an item name of an output of the second source model, or when it is determined that dimensional information of a unit of an input of the first source model and dimensional information of a unit of an output of the second source model are similar.

2. In claim 1, The model generation unit includes a model determination unit, the model determination unit displays, on a GUI screen, an input node indicating an input of the target model, and a model combination diagram indicating the first source model and the second source model; the model combination diagram connects an input node of the target model to an input node indicating an input of the corresponding first source model or the corresponding second source model with an edge, and connects an output node indicating an output of the second source model to an input node indicating an input of the corresponding first source model with an edge; a model determination unit that, when a connection via an edge of the model connection diagram is modified on the GUI screen, modifies the connection between the input of the target model and the input of the first source model or the second source model, or the connection between the output of the second source model and the input of the first source model, in accordance with the modified connection.

3. In claim 1, The source model stored in the source model database includes an equation or an inequality having an equality condition, A model generation system in which a source model representing the equation or inequality is defined with one parameter of the equation or inequality as an output and the other parameter as an input.

4. In claim 1, The source model stored in the source model database includes an inequality, A model generation system in which a source model representing the inequality is defined as having as output a Boolean value indicating whether the inequality is satisfied or not, and all parameters as input.

5. A model generation system for generating a target model, comprising: a source model database for storing source models; a model generation unit for generating the target model from a source model retrieved from the source model database; A machine learning unit that performs machine learning, the model generation unit includes a database search unit that searches for a first source model that includes an output of the target model as its output and a second source model that includes an input of the target model as its input, a combination determination unit, and a model determination unit; the model determination unit displays on a GUI screen a model combination diagram including an input node indicating an input of the target model and one or more search source models including the first source model and the second source model searched by the database search unit, the model combination diagram connecting corresponding input nodes with each other or connecting an output node of a search source model with a corresponding input node of another search source model with an edge; the model combination diagram connects an input node of the target model to an input node indicating an input of the corresponding first source model or the corresponding second source model with an edge, and connects an output node indicating an output of the second source model to an input node indicating an input of the corresponding first source model with an edge; the combination determination unit combines the input of the other search source model with the output of the certain search source model when the input of the other search source model can be associated with the output of the certain search source model; the combination determination unit determines that correspondence is possible when an input item name of the certain search source model matches an output item name of the other search source model, when an input item name of the certain search source model and an output item name of the other search source model are synonymous, when there is a combination history between an input item name of the certain search source model and an output item name of the other search source model, or when it is determined that unit dimensional information of the input of the certain search source model and unit dimensional information of the output of the other search source model are similar, when a connection by an edge of the model connection diagram is corrected on the GUI screen, the model determination unit corrects a connection between an input of the target model and an input of the first source model or an input of the second source model, or a connection between an output of the second source model and an input of the first source model, in accordance with the corrected connection; the model determination unit determines a combined model obtained by modifying a combination of an input of the target model, the search source model, and the unlearned model in accordance with the modified connection when a connection of the model combination diagram is modified so as to add an unlearned model on the GUI screen; The machine learning unit is a model generation system that trains the combined model using learning data corresponding to the input and output of the target model.

6. A model generation system for generating a target model, comprising: a source model database for storing source models; a model generation unit for generating the target model from a source model retrieved from the source model database; A machine learning unit that performs machine learning, the model generation unit includes a database search unit that searches for a first source model that includes an output of the target model as its output and a second source model that includes an input of the target model as its input, and a model determination unit; the model determination unit displays on a GUI screen a model combination diagram including an input node indicating an input of the target model and one or more search source models searched by the database search unit, in which corresponding input nodes are connected to each other or an output node of a search source model is connected to a corresponding input node of another search source model by an edge; the model determination unit determines a combined model obtained by modifying a combination of an input of the target model, the search source model, and the unlearned model in accordance with the modified connection when a connection of the model combination diagram is modified so as to add an unlearned model on the GUI screen; The machine learning unit trains the combined model using learning data corresponding to the input and output of the target model; the machine learning unit sets a standard model, which is a machine learning model that receives the input of the target model as an input and receives the output of the target model as an output, and trains the standard model using the learning data; The model determination unit is a model generation system that selects either the combined model or the standard model as the target model.

7. In claim 5 or 6, A model generation system in which the machine learning unit sets the learning rate of the search source models included in the combined model to 0 when training the combined model.

8. In claim 5 or 6, The source model stored in the source model database includes an equation or an inequality having an equality condition, A model generation system in which a source model representing the equation or inequality is defined with one parameter of the equation or inequality as an output and the other parameter as an input.

9. In claim 5 or 6, The source model stored in the source model database includes an inequality, A model generation system in which a source model representing the inequality is defined as having as output a Boolean value indicating whether the inequality is satisfied or not, and all parameters as input.

10. A model generation method for generating a target model using a model generation system, comprising: The model generation system includes a source model database for storing a source model, and a model generation unit for generating the target model by using a source model searched from the source model database, The model generation unit searching for a first source model that includes as its output the output of the target model and a second source model that includes as its input the input of the target model; combining the input of the first source model with the output of the second source model when the input of the first source model and the output of the second source model can be associated with each other; The source model stored in the source model database includes a trained machine learning model, a model generation method in which the model generation unit determines that correspondence is possible when an item name of an input of the first source model matches an item name of an output of the second source model, when an item name of an input of the first source model and an item name of an output of the second source model are synonyms, when there is a binding history between an item name of an input of the first source model and an item name of an output of the second source model, or when it is determined that dimensional information of a unit of an input of the first source model and dimensional information of a unit of an output of the second source model are similar.

11. In claim 10, The model generation unit displaying an input node indicating an input of the target model, and a model combination diagram indicating the first source model and the second source model on a GUI screen, the model combination diagram connecting an input node of the target model to an input node indicating an input of the corresponding first source model or the second source model with an edge, and connecting an output node indicating an output of the second source model to an input node indicating an input of the corresponding first source model with an edge; A model generation method in which, when a connection via an edge of the model connection diagram is modified on the GUI screen, the connection between the input of the target model and the input of the first source model or the second source model, or the connection between the output of the second source model and the input of the first source model is modified in accordance with the modified connection.

Images