JP2021089655A - Learning model construction device, method for constructing learning model, and computer program - Google Patents

Abstract

To provide a technique of generating a mathematical model which can be predicted more accurately than a production system.SOLUTION: A learning model construction device includes: a first training information acquisition unit for acquiring first training information showing construction input data input in a production system and a rule that a fire occurs when the construction input data is input; a second training information acquisition unit for acquiring second training information showing the rule that a fire occurs when the construction input data is input and an output of the production system when the construction input data is input; a first layer learning unit as a first neural network for learning with the first training information; a second layer learning unit as a second neural network for learning with the second training information; and a stacking unit for connecting the first neural network and the second neural network to each other so that output of a first-learned model obtained by the first layer learning unit is input into a second-learned model obtained by the second layer learning unit.SELECTED DRAWING: Figure 1

Description

The present invention relates to a learning model construction device, a learning model construction method, and a computer program.

There is a mathematical model called a production system that predicts the state of the prediction target according to one or a plurality of preset rules based on the input information. In the production system, the preset rules are the rules assumed in advance by the designer of the production system. More specifically, the preset rule is a rule set by the designer of the production system based on the information input in advance by the designer of the production system. Examples of the use of such a production system include use in natural language processing. In such a case, the production system determines the meaning of the input sentence by determining whether or not there is a predetermined specific term in the input sentence.

Japanese Unexamined Patent Publication No. 05-165886

By the way, as described above, the rules in the production system are set for the input assumed by the designer. Therefore, there is a problem that the production system cannot output an appropriate prediction result when an input different from the designer's assumption is made. For example, in the case of being used for the above-mentioned natural language processing, when the input sentence is a sentence that conveys the meaning in a roundabout expression beyond the designer's expectation, the production system has the meaning indicated by the input sentence. May not be predictable.

In view of the above circumstances, an object of the present invention is to provide a technique for generating a mathematical model having higher prediction accuracy than a production system.

One aspect of the present invention is the construction input data input to the production system to be converted into the neural network learned by the machine learning method, and the production system when the construction input data is input. The first training information acquisition unit that acquires the first training information indicating the information indicating the ignition rule, the information indicating the rule that ignites in the production system when the construction input data is input, and the construction Description of the second training information acquisition unit that acquires the second training information indicating the production output that is the output of the production system when the input data is input, and the training data for the construction input data of the first training information. The first layer learning unit, which is a first neural network that is used as a variable and uses the information indicating the rule of the first training information as the objective variable of the training data to be learned by the machine learning method, and the rule of the second training information. The second layer learning unit, which is a second neural network that learns by a machine learning method that uses the information indicating the above as an explanatory variable of the training data and uses the production output of the second training data as the objective variable of the training data, and the above. The first trained model, which is a trained model obtained by the first layer learning unit, is a trained model in which the output of the first trained model is input to the second trained model, which is the trained model obtained by the second layer learning unit. 3 Learning including a stacking unit that connects the first neural network and the second neural network so that the trained model is expressed by using the first neural network and the second neural network. It is a model building device.

One aspect of the present invention is the above-mentioned learning model construction device, in which the data input to the third trained model is used as an explanatory variable of the training data, and the third trained when the data is input. It further includes a third layer learning unit, which is a third neural network that learns by a machine learning method that uses the output of the model as an objective variable of training data.

One aspect of the present invention is the training model construction device, and the third neural network is added to additional training data which is training data additionally input after the generation of the third trained model. Learn more based on.

One aspect of the present invention is the above-mentioned learning model construction device, in which the degree of agreement between the production output when the explanatory variables are input to the production system and the teacher data determined by the user with respect to the information of the explanatory variables is. The first explanatory variable condition is a condition that is lower than the standard matching degree, which is a predetermined matching degree, and the first purpose is the condition that the teacher data is determined by the user for the explanatory variable that satisfies the first explanatory variable condition. The second explanatory variable condition is a variable condition in which the degree of agreement between the production output when the explanatory variable is input to the production system and the teacher data determined by the user with respect to the information of the explanatory variable is equal to or higher than the reference degree of agreement. As a second objective variable condition, the condition that the teacher data is determined by the user for the explanatory variable satisfying the second explanatory variable condition is set as the second objective variable condition, the explanatory variable satisfies the first explanatory variable condition, and the objective variable is the first objective variable. The additional training data includes training data satisfying one objective variable condition and training data in which the explanatory variable satisfies the second explanatory variable condition and the objective variable satisfies the second objective variable condition.

One aspect of the present invention is the above-mentioned learning model construction device, and the second neural network further learns based on production input / output training information which is a set of data of a construction input data and a production output. To do.

One aspect of the present invention is the construction input data input to the production system to be converted into the neural network learned by the machine learning method, and the production system when the construction input data is input. The first training information acquisition step for acquiring the first training information indicating the information indicating the firing rule, the information indicating the rule indicating the firing rule in the production system when the construction input data is input, and the construction Explanation of the second training information acquisition step for acquiring the second training information indicating the production output which is the output of the production system when the input data is input, and the training data for the construction input data of the first training information. The first layer learning step that the first neural network learns by the machine learning method that uses the information indicating the rule of the first training information as a variable and uses the information indicating the rule of the first training data as the objective variable of the training data, and the rule of the second training information. The second layer learning step in which the second neural network learns by a machine learning method using the indicated information as an explanatory variable of the training data and the production output of the second training data as the objective variable of the training data, and the first The output of the first trained model, which is the trained model obtained in the layer learning step, is the trained model that is input to the second trained model, which is the trained model obtained in the second layer learning step. 3 Learning having a stacking step for connecting the first neural network and the second neural network so as to represent the trained model using the first neural network and the second neural network. This is a model construction method.

One aspect of the present invention is a computer program for operating a computer as the above-mentioned learning model construction device.

INDUSTRIAL APPLICABILITY According to the present invention, it is possible to generate a mathematical model with higher prediction accuracy than a production system.

Explanatory drawing explaining the outline of operation of the learning model construction apparatus 1 of embodiment. The figure which shows an example of the hardware composition of the learning model construction apparatus 1 of embodiment. The figure which shows an example of the functional structure of the control part 10 in embodiment. The flowchart which shows an example of the flow of the process executed by the learning model construction apparatus 1 of embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is an explanatory diagram illustrating an outline of the operation of the learning model construction device 1 of the embodiment.
The learning model construction device 1 converts the production system 9 to be converted into a neural network to be learned by a machine learning method based on the construction input data, the production output, and the firing rule information. The construction input data is information input to the production system 9. The production output is the output of the production system 9 when the construction input data is input. The ignition rule information is information indicating a rule (hereinafter referred to as “ignition rule”) for firing in the production system 9 when input data for construction is input.

Converting the production system 9 to a neural network means generating a neural network that represents the model to be constructed. The built model outputs information in which the difference from the information output by the production system 9 when the expected input data is input to the production system 9 is smaller than the predetermined difference when the expected input data is input. It is a trained model to do. As an example of a predetermined difference, for example, when the information to be output is a vector, the similarity between the vectors based on the Cos similarity estimation method is between the vector output by the production system 9 and the vector output by the model to be constructed. Is within the specified range. The constructed model is a trained model that outputs a prediction result based on the learning result when information other than the expected input data is input. The expected input data is an input that the designer of the production system 9 assumes in advance that the production system 9 can make an appropriate prediction when it is input to the production system 9.

The learning model construction device 1 generates the first trained model based on the first training information. The first training information is information indicating the construction input data and the ignition rule indicated by the ignition rule information. That is, the first training information is a set of information of the construction input data and the firing rule information. The first training information is training data for generating the first trained model.

The first trained model shows a rule that is expected to fire in the production system 9 when the information indicated by the input expected input data is input to the production system 9 when the expected input data is input. Output information.

More specifically, the training model construction device 1 generates the first trained model by a machine learning method using the construction input data as an explanatory variable of the training data and the firing rule information as the objective variable of the training data.

The learning model construction device 1 generates a second trained model based on the second training information. The first training information is information indicating the ignition rule and the production output indicated by the ignition rule information. That is, the second training information is the information of the set of the firing rule information and the production output. The second training information is training data for generating the second trained model.

The second trained model is a production that is expected to be output by the production system 9 when the expected ignition rule information is input and the ignition rule indicated by the input expected ignition rule information is ignited in the production system 9. Output the output. The expected firing rule information is information indicating a rule for firing in the production system 9 when the expected input data is input to the production system 9.

More specifically, the learning model construction device 1 generates a second trained model by a machine learning method using the firing rule information as an explanatory variable of the training data and the production output as the objective variable of the training data.

The learning model construction device 1 generates a model to be constructed based on the first trained model and the second trained model. More specifically, the learning model construction device 1 performs a process including a connection process (hereinafter referred to as "finishing process") so that the output of the first trained model is input to the second trained model. , Generate a model to be constructed. The connection process is a process of connecting a neural network representing the first trained model and a neural network representing the second trained model. The finishing process may include, for example, a so-called distillation process in addition to the connection process. Specifically, in the finishing process, the distillation process may be executed after the connection process is executed.

FIG. 2 is a diagram showing an example of the hardware configuration of the learning model construction device 1 of the embodiment. The learning model construction device 1 includes a CPU (Central Processing Unit), a memory, an auxiliary storage device, and the like connected by a bus, and by executing a program, a control unit 10, an interface unit 11, a storage unit 12, an input unit 13, and the like. It functions as a device including the output unit 14.

More specifically, in the learning model construction device 1, the processor 91 reads a program stored in the storage unit 12, and stores the read program in the memory 92. When the processor 91 executes the program stored in the memory 92, the learning model construction device 1 functions as a device including the control unit 10, the interface unit 11, the storage unit 12, the input unit 13, and the output unit 14.

All or a part of each function of the learning model construction device 1 may be realized by using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a flexible disk, a magneto-optical disk, a portable medium such as a ROM or a CD-ROM, or a storage device such as a hard disk built in a computer system. The program may be transmitted over a telecommunication line.

The interface unit 11 includes a communication interface for connecting the own device to an external device such as a production system 9. The interface unit 11 communicates with, for example, the production system 9 via, for example, a network. The interface unit 11 acquires construction input data, firing rules, and production output from the production system 9 by communicating with the production system 9. One of the external devices may be, for example, an external storage device such as a USB (Universal Serial Bus) memory. The interface unit 11 outputs a program representing the model to be constructed generated by the learning model construction device 1 to an external storage device such as a USB (Universal Serial Bus) memory.

The storage unit 12 is configured by using a storage device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 12 stores various information related to the learning model construction device 1. The storage unit 12 stores, for example, the construction input data, the firing rule, and the production output acquired from the production system 9.

The input unit 13 includes an input device such as a mouse, a keyboard, and a touch panel. The input unit 13 may be configured as an interface for connecting these input devices to its own device. The input unit 13 receives input of various information to its own device. The various information is, for example, information indicating an instruction to start learning to the learning model construction device 1.

The output unit 14 includes display devices such as a CRT (Cathode Ray Tube) display, a liquid crystal display, and an organic EL (Electro-Luminescence) display. The output unit 14 may be configured as an interface for connecting these display devices to its own device. The output unit 14 displays information about the own device.

FIG. 3 is a diagram showing an example of the functional configuration of the control unit 10 in the embodiment. The control unit 10 includes a first training information acquisition unit 111, a second training information acquisition unit 112, a first layer learning unit 113, a second layer learning unit 114, a stacking unit 115, a third layer learning unit 116, and an interface control unit 117. To be equipped.

The first training information acquisition unit 111 acquires the first training information via the interface unit 11. The second training information acquisition unit 112 acquires the second training information via the interface unit 11.

The first layer learning unit 113 uses the input data for constructing the first training information as the explanatory variable of the training data, and the firing rule information of the first training information is used as the objective variable of the training data. It is a network. Hereinafter, the neural network of the first layer learning unit 113 is referred to as a first neural network. The trained model of the learning model represented by the first neural network is the first trained model. The first layer learning unit 113 generates the first trained model by learning based on the first training information. That is, the first trained model is a trained model obtained by the first layer learning unit.

The second layer learning unit 114 is a neural network that learns by a machine learning method that uses the firing rule information of the second training information as an explanatory variable of the training data and the production output of the second training information as the objective variable of the training data. is there. Hereinafter, the neural network of the second layer learning unit 114 is referred to as a second neural network. The trained model of the learning model represented by the second neural network is the second trained model. The second layer learning unit 114 generates a second trained model by learning based on the second training information. That is, the second trained model is a trained model obtained by the second layer learning unit.

The stacking unit 115 connects the first neural network and the second neural network so as to represent the third trained model using the first neural network and the second neural network. That is, the stacking unit 115 causes the output of the first neural network to be input to the second neural network. The third trained model is a trained model including the first trained model and the second trained model, and the output of the first trained model is input to the second trained model. The third trained model is an example of the model to be constructed. The process executed by the stacking unit 115 is an example of the connection process.

Hereinafter, a neural network including a first neural network and a second neural network and whose output of the first neural network is an input of the second neural network is referred to as a stacked neural network. The trained model of the training model represented by the stacked neural network is an example of the model to be constructed.

The third layer learning unit 116 learns by a machine learning method that uses the inheritance data as an explanatory variable of the training data and uses the output of the third trained model when the inheritance data is input as the objective variable of the training data. It is a neural network. The inheritance data is the data input to the third trained model. Hereinafter, the neural network of the third layer learning unit 116 is referred to as a third neural network. The third layer learning unit 116 generates the inherited trained model by learning based on the third training information.

The inherited trained model is a trained model of the learning model represented by the third neural network. The third training information is information indicating the inheritance data and the output of the third trained model when the inheritance data is input. That is, the third training information is a set of information of the inheritance data and the output of the third trained model when the inheritance data is input. The process executed by the third layer learning unit 116 is a so-called distillation process. The process executed by the third layer learning unit 116 may be performed using so-called GAN (Generative Adversarial Networks). The inherited trained model is an example of the constructed model.

The interface control unit 117 controls the operation of the interface unit 11. The interface control unit 117 controls the operation of the interface unit 11, for example, and outputs the model to be constructed to the external device.

FIG. 4 is a flowchart showing an example of the flow of processing executed by the learning model building apparatus 1 of the embodiment.
The first training information acquisition unit 111 acquires the first training information (step S101). Next, the second training information acquisition unit 112 acquires the second training information (step S102). Next, the first layer learning unit 113 uses a machine learning method in which the input data for constructing the first training information is used as the explanatory variable of the training data and the firing rule information of the first training information is used as the objective variable of the training data. Then, the first trained model is generated (step S103). Next, the second layer learning unit 114 uses a machine learning method that uses the firing rule information of the second training information as the explanatory variable of the training data and the production output of the second training information as the objective variable of the training data. A second trained model is generated (step S104).

Next, the stacking unit 115 connects the first neural network and the second neural network so as to represent the third trained model using the first neural network and the second neural network (step). S105). Next, the third layer learning unit 116 generates an inherited trained model by learning based on the third training information (step S106).

The learning model construction device 1 of the embodiment configured in this way generates a model to be constructed by machine learning using the input data for construction, the firing rule information, and the production output. Since the trained model generated by the machine learning method is a mathematical model with extrapolation function, the trained model makes highly accurate predictions even when information that was not learned in the learning stage is input. The result can be output. In this way, the learning model construction device 1 can generate a mathematical model with higher prediction accuracy than the production system.

Further, the learning model construction device 1 of the embodiment configured in this way learns by using the ignition rule information in addition to the construction input data and the production output. Therefore, it is possible to generate a mathematical model with higher prediction accuracy than machine learning using only construction input data and production output.

(Modification example)
The second neural network may be learned not only based on the second training information but also based on the production input / output training information. The production input / output training information is a set of data of the construction input data and the production output.

In such a case, the second neural network learns while estimating the probability that the relationship between the explanatory variable and the objective variable indicated by the second training information is correct based on the relationship between the construction input data and the production output. Can be done. Therefore, the second trained model as a result of learning using the second training information and the production input / output training information outputs a more accurate result than the second trained model trained using only the second training information. can do.

The learning model building apparatus 1 may further perform additional learning after step S106. In such a case, specifically, the third neural network of the third layer learning unit 116 further learns based on the additionally input training data (hereinafter referred to as “additional training data”). The following step S106 means after the generation of the third trained model.

In such a case, since the inherited trained model is further trained from the time of step S106, the prediction accuracy is higher than that of the trained model generated in the process of step S106. The trained model as a result of learning by the inherited trained model using the additional training data is also an example of the constructed model.

The additional training data includes, for example, training data satisfying the following first explanatory variable condition and first objective variable condition (hereinafter referred to as “mismatch data”), and training data satisfying the second explanatory variable condition and second objective variable condition. (Hereinafter referred to as "matching data") and may be included.

The first explanatory variable condition is that the explanatory variable has a predetermined degree of coincidence between the production output when the explanatory variable is input to the production system 9 and the teacher data determined by the user with respect to the information of the explanatory variable. The condition is that it is lower than the "standard degree of agreement").

The first objective variable condition is a condition that the teacher data is determined by the user for the explanatory variable satisfying the first explanatory variable condition.

The second explanatory variable condition is that the degree of agreement between the production output when the explanatory variable is input to the production system 9 and the teacher data determined by the user with respect to the information of the explanatory variable is equal to or greater than the reference degree. It is a condition.

The second objective variable condition is a condition that the teacher data is determined by the user for the explanatory variable satisfying the second explanatory variable condition.

In such a case, the inherited trained model is trained based on more accurate information than when trained based only on the result of the production system 9. Therefore, the trained model trained using the additional training data including the match data and the mismatch data has higher prediction accuracy than the trained model generated in the process of step S106. The trained model as a result of learning by the inherited trained model using the additional training data including the match data and the mismatch data is also an example of the constructed model.

When learning using additional training data including match data and mismatch data, for example, even if training is performed using a loss function whose value increases when the prediction of mismatch data fails in the prediction at the time of training. Good.

The first trained model may be a trained model provided with a discriminator for determining whether or not each rule fires for each rule of the production system 9. The first trained model may be one trained model that outputs information indicating which of the rules of the production system 9 has fired.

It should be noted that steps S101 and S102 do not necessarily have to be executed in the order shown in FIG. Step S101 may be executed at any timing as long as it is executed before the execution of step S103. Step S102 may be executed at any timing as long as it is executed before the execution of step S104. The processes of steps S103 and S104 do not necessarily have to be executed in the order shown in FIG. The processing of step S103 and step S104 may be executed at any timing as long as it is executed before the processing of step S105 is executed. For example, the process of step S103 may be executed before the execution of step S105, and the process of step S104 may be executed before the execution of the process of step S103.

The learning is to appropriately adjust the parameters of the learning model. The trained model is a trained model at the time when the end condition is satisfied. The end condition may be any condition as long as it is a condition related to the end of learning. The end condition may be, for example, a condition that learning with a predetermined number of training data has been executed, or a condition that the amount of change in parameters due to learning is less than a predetermined magnitude.

The neural network may be any neural network as long as the accuracy of the output result is equal to or higher than a predetermined accuracy. The neural network may be, for example, a forward propagation type neural network or an RBF (Radial Basis Function) network. The neural network may be a convolutional neural network, a recurrent neural network, or a stochastic neural network. The neural network may be a spiking neural network or a complex neural network.

The parameters of the learning model may be adjusted by any machine learning algorithm according to the type of neural network, as long as it is a learning algorithm with high accuracy of the output result. For example, the parameters of the learning model may be adjusted by the algorithm of the error back propagation method if it is a neural network other than the forward propagation type neural network.

The learning model construction device 1 may be implemented by using a plurality of information processing devices connected so as to be able to communicate via a network. In this case, each functional unit included in the learning model construction device 1 may be distributed and mounted in a plurality of information processing devices. For example, the third layer learning unit 116 may be mounted on an information processing device different from other functional units included in the control unit 10.

Although the embodiments of the present invention have been described in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and includes designs and the like within a range that does not deviate from the gist of the present invention.

1 ... Learning model construction device, 10 ... Control unit, 11 ... Interface unit, 12 ... Storage unit, 13 ... Input unit, 14 ... Output unit, 111 ... 1st training information acquisition unit, 112 ... 2nd training information acquisition unit, 113 ... 1st layer learning unit, 114 ... 2nd layer learning unit, 115 ... Stacking unit, 116 ... 3rd layer learning unit, 117 ... Interface control unit

Claims (7)

Construction input data input to the production system to be converted into a neural network learned by a machine learning method, and information indicating a rule that fires in the production system when the construction input data is input. The first training information acquisition unit that acquires the first training information indicating
Second training information indicating a rule indicating a rule to ignite in the production system when the construction input data is input and a production output which is an output of the production system when the construction input data is input. The second training information acquisition department to acquire
In a first neural network that learns by a machine learning method that uses the input data for constructing the first training information as an explanatory variable of the training data and uses the information indicating the rule of the first training information as the objective variable of the training data. With a certain first layer learning department
A second neural network that learns by a machine learning method that uses information indicating the rules of the second training information as an explanatory variable for training data and uses the production output of the second training information as an objective variable for training data. Two-layer learning department and
It is a trained model in which the output of the first trained model, which is the trained model obtained by the first layer learning unit, is input to the second trained model, which is the trained model obtained by the second layer learning unit. A stacking unit that connects the first neural network and the second neural network so as to express the third trained model using the first neural network and the second neural network, and
A learning model construction device equipped with. Learning by a machine learning method in which the data input to the third trained model is used as an explanatory variable of the training data and the output of the third trained model when the data is input is used as the objective variable of the training data. Third layer learning unit, which is the third neural network
The learning model construction apparatus according to claim 1. After the generation of the third trained model, the third neural network further learns based on the additional training data which is the additional training data input.
The learning model building apparatus according to claim 2. The first condition is that the degree of agreement between the production output when the explanatory variable is input to the production system and the teacher data determined by the user with respect to the information of the explanatory variable is lower than the reference degree of agreement, which is a predetermined degree of agreement. Production when the explanatory variable is input to the production system, with the condition that the explanatory variable condition is the teacher data determined by the user for the explanatory variable satisfying the first explanatory variable condition as the first objective variable condition. The condition that the degree of matching between the output and the information of the explanatory variable with the teacher data determined by the user is equal to or higher than the reference degree of matching is set as the second explanatory variable condition, and the user with respect to the explanatory variable satisfying the second explanatory variable condition As the second objective variable condition, the training data that the explanatory variable satisfies the first explanatory variable condition and the objective variable satisfies the first objective variable condition, and the explanatory variable is the second objective variable condition. The additional training data includes training data that satisfies the variable condition and the objective variable satisfies the second objective variable condition.
The learning model building apparatus according to claim 3. The second neural network further learns based on the production input / output training information, which is a set of data of the construction input data and the production output.
The learning model construction device according to any one of claims 1 to 4.
Construction input data input to the production system to be converted into a neural network learned by a machine learning method, and information indicating a rule that fires in the production system when the construction input data is input. The first training information acquisition step to acquire the first training information indicating
Second training information indicating a rule indicating a rule to ignite in the production system when the construction input data is input and a production output which is an output of the production system when the construction input data is input. The second training information acquisition step to acquire
The first neural network learns by a machine learning method in which the input data for constructing the first training information is used as an explanatory variable of the training data and the information indicating the rule of the first training information is used as the objective variable of the training data. First layer learning step and
The second neural network learns by a machine learning method in which the information indicating the rule of the second training information is used as an explanatory variable of the training data and the production output of the second training information is used as the objective variable of the training data. Layer learning steps and
A trained model in which the output of the first trained model, which is the trained model obtained in the first layer learning step, is input to the second trained model, which is the trained model obtained in the second layer learning step. A stacking step for connecting the first neural network and the second neural network so as to represent the third trained model using the first neural network and the second neural network.
How to build a learning model with. A computer program for operating a computer as the learning model construction device according to any one of claims 1 to 5.

Images