JPWO2019064461A1 - Learning network generation device and learning network generation program - Google Patents

Abstract

It is an object of the present invention to provide a learning network generation device capable of automatically generating a learning network having a structure suitable for an object. The learning network generation device generates an initial individual generation unit that generates data defining an individual of the learning network, and generates a program code corresponding to the individual based on the data defining the individual of the learning network generated by the initial individual generation unit. Code generation unit, a learning execution unit that applies evolutionary learning by applying learning data to program codes corresponding to individuals generated by the code generation unit, and a learning execution unit that performs evolutionary learning An evaluation / selection unit that evaluates the individual and determines whether to leave the individual according to the result of the evaluation.

Description

  The present invention relates to a program for generating a learning network that performs machine learning by so-called deep learning (deep learning).

  In recent years, in the field of artificial intelligence, "machine learning", in which a machine (computer) automatically derives decision rules from learning data by inputting a lot of learning data without describing rules, has attracted attention. I am collecting. Deep learning, which is a type of machine learning, is increasingly applied to fields such as image recognition and speech recognition. It is said that deep learning is characterized in that a machine (computer) automatically learns features to be extracted from learning data. In other words, it is said that in deep learning, it is not necessary to teach what to focus on, and it is possible to automatically learn what features can be used to correctly identify an object. Various devices using such deep learning have been proposed (for example, see Patent Document 1).

JP, 2017-111660

  However, in order for a computer to actually perform deep learning, it is necessary to define a learning network in advance. Here, the learning network is an algorithm using nerve cells (neurons) of the brain of an organism as a model and includes a plurality of layers. Each layer is composed of one or more nodes (also referred to as units; hereinafter, referred to as units to distinguish them from nodes in a tree structure described later). The learning network adjusts the weight (coefficient) of the input to each node by repeatedly performing machine learning, thereby increasing the accuracy (correct answer rate) at which a correct answer can be output based on the input.

  Meanwhile, as shown in FIG. 1, a learning network is often represented by a multilayer perceptron model including an input layer, one or more intermediate layers, and an output layer. In this learning network model, each layer has one or more nodes, and one or more input values are given to each unit from the previous layer (for the input layer, from the outside). The number of units included in each layer and the connection relation of units between layers are various. The input values for the units are each multiplied by a weighting factor, and the sum of the values obtained by multiplying the input values by the weighting factor is the argument of the activation function defined for each unit. The output of the activation function is an output value to the next layer.

  As described above, the structure of the learning network is defined by the number of units constituting each layer, the activation function in each unit, the connection relationship between layers, and the like. When an input is given to the defined learning network, Learning is performed by correcting the weight coefficient applied to the input to each unit so that the output approaches the correct answer.

  Therefore, if the structure of the learning network is inappropriate, even if machine learning is repeated, the correct answer rate cannot be increased. As described above, the structure of the learning network is an important element in enhancing the effect of deep learning, but the structure of the appropriate learning network depends on the target (for example, whether the target is character identification or image identification, etc.). )different. For this reason, conventionally, a designer has basically found out and defined a structure of a learning network that seems appropriate for an object by repeating trial and error by hand.

  The present invention has been made to solve the above problems, and has as its object to provide a learning network generation device and a learning network generation program capable of automatically generating a learning network having a structure suitable for an object. I do.

  In order to solve the above problems, a learning network generation device according to the present invention includes an initial individual generation unit that generates data defining an individual of a learning network, and data that defines an individual of the learning network generated by the initial individual generation unit. A code generation unit that generates a program code corresponding to the individual, a learning execution unit that executes evolutionary learning by applying learning data to the program code corresponding to the individual generated by the code generation unit, An evaluation / selection unit that evaluates the learned individual that has performed the evolutionary learning in the learning execution unit and determines whether to leave the individual according to the evaluation result. At this time, the initial individual generation unit may generate data representing the structure of the learning network.

  With such a configuration, a learning network having a structure suitable for an object can be automatically generated.

  Also, in the present invention, the initial individual generation unit converts data representing the structure of the learning network into a layer node corresponding to one layer in the learning network, and at least a layer node indicating a connection relationship between the one layer and the previous layer; The data may be output as tree-structured data including a parameter node indicating a parameter related to a layer structure corresponding to the layer node.

  In this way, the structure of the individual of the learning network can be expressed in a format that is easy to handle.

  According to the present invention, when the learning network has a structure in which a plurality of layers each having one or more units are connected, data representing the structure of the learning network includes the number of units constituting each layer, the activation in each unit. It is preferable to include a function and a connection relationship between layers.

  In the present invention, the code generation unit may generate a program code corresponding to an individual based on data representing the structure of the learning network generated by the initial individual generation unit and template data of a common part independent of the individual.

  In the present invention, the learning network generation device further includes an evolution execution unit that generates an individual of the next generation by performing an evolution process on the individuals selected by the evaluation / selection unit, and the code generation unit includes an evolution execution unit. It is preferable to generate a program code corresponding to the generated next-generation individual. In this way, a new generation of individuals that evolved from superior individuals is automatically generated by multiplying superior individuals or causing mutations in superior individuals. An excellent individual can be output by performing the selection.

  The learning network generation program according to the present invention may cause a computer to function as any one of the learning network generation devices described above.

  A learning network generating program according to another example of the present invention includes: an initial individual generating step of generating data defining an individual of a learning network; and data defining an individual of the learning network generated in the initial individual generating step. A code generation step of generating a program code corresponding to the individual, and a learning execution step of executing evolutionary learning by applying learning data to the program code corresponding to the individual generated in the code generation step. And evaluating a learned individual that has performed the evolutionary learning in the learning execution step, and determining whether to leave the individual according to the evaluation result.

FIG. 3 is a schematic diagram showing an example of expression of a learning network by a multilayer perceptron model. FIG. 3 is a schematic diagram illustrating an example of expression using a tree structure in which a learning network is handled by a learning network generation device. It is a schematic diagram which shows the example of a structure of a learning network production | generation apparatus. FIG. 2 is a functional block diagram of the learning network generation device 1. It is a flowchart which shows the procedure of a learning network generation process. It is a schematic diagram explaining the crossing by the one-point method.

  Hereinafter, a learning network generation device 1 according to an embodiment of the present invention will be described with reference to the drawings.

[Data Structure of Learning Network Handled by Learning Network Generation Apparatus 1]
The learning network generation device 1 according to the present embodiment defines each learning network (individual) with a tree structure. FIG. 2 shows an example of an individual defined by a tree structure. In the tree structure illustrated in FIG. 2, the vertical direction of the drawing represents a hierarchy (layer), and a layer drawn higher in the drawing is a higher layer. In the same layer, a node drawn on the right end is called a layer node LN, and represents an attribute of each layer in the learning network. In the same layer, nodes drawn at positions other than the right end are called parameter nodes PN, and define parameters of a layer node LN one layer higher. For example, the layer node LN of the second hierarchy L2 is "PreviousDense", and the value of each parameter node PN belonging to the third hierarchy L3 is set as a parameter in "PreviousDense". In addition, as described later, the significance of the value of the parameter node PN is defined for each type of the layer node LN. Further, the existence of a layer node LN having no parameter node PN, such as 'Flatten' which is a layer node LN of the third hierarchy L3 in FIG. 2, is also allowed.

  In the learning network generation device 1, a learning network represented by such a tree structure is described by character string data. A character string describing a tree structure starts with “{”. Then, a node number (eg, “0:”, “1:”, etc.) in the tree structure, a character string indicating the type of the layer node LN (eg, “HyperParameter”, “PreviousDense”, etc.) or the immediately preceding layer node LN The structure of the learning network is described by repeating the values of the continuous parameter nodes PN and delimiters (for example, “,”), and the end of the description of the tree structure is specified by describing “}” at the end. For example, the tree-structured learning network shown in FIG. 2 is {0: 'HyperParameter', 1: 0.425, 2: 0.423, 3: 'PreviousDense', 4: 37, 5: 3, 6: 0.52, 7: ' Flatten ', 8:' FollowingDense ', 9: 136, 10: 12, 11: 0.571, 12:' LastDense ', 13: 1, 14: 0, 15: 0.053}.

  The correspondence between the learning network represented by the tree structure and the learning network drawn by the multilayer perceptron model is as follows.

  In the tree structure expression, in all the individuals, the layer node LN of the highest hierarchy L1 is “HyperParameter”, and two parameters (“LEARNING_RATE” and “MOMENTUM”) for the “HyperParameter” are the parameters of the second hierarchy L2. 'HyperParameter' defined as a value of the node PN is a parameter related to learning of an individual, and is not directly related to a structure of a learning network represented by a multilayer perceptron model. Note that the above “LEARNING_RATE” and “MOMENTUM” are merely examples, and the number and type of parameter nodes PN connected to “HyperParameter” are arbitrary.

  Further, in the learning network generating device 1, the input layer and the input to the input layer in the multilayer perceptron model are common to all individuals and do not appear in the representation of the tree structure.

  The names of the layer nodes LN in the second and lower layers L2 and lower indicate the connection method with the previous layer in the multilayer perceptron model. That is, for example, the layer node LN of the second layer L2 indicates a connection method between the input layer and the first intermediate layer. The parameter node PN connected to the layer node LN in the second and lower layers L2 and below is defined according to the number of units of the layer corresponding to the layer node LN, the type of the activation function, and other types of the layer node LN. Parameters (eg, parameters applied to the activation function). Examples of the activation function include a step function, a sigmoid function, a ReLU function, an identity function, a softmax function, and a hyperbolic tangent (tanh) function.

  As described above, the learning network generation device 1 of the present embodiment expresses the structure of each layer in the multilayer perceptron model of the individual of the learning network and the connection relation thereof by the tree structure. The weight coefficient multiplied by the output value passed between units is not specified in the expression of the tree structure. The learning network generation device 1 sets a random value as an initial value of the weight coefficient before learning when the individual of the learning network performs learning.

  Hereinafter, the configuration of the learning network generation device 1 that generates the learning network represented by the tree structure as described above will be described in detail.

[Hardware Configuration of Learning Network Generation Apparatus 1]
FIG. 3 shows a configuration example of the learning network generation device 1 according to the embodiment of the present invention. The learning network generation device 1 generates a learning network by a genetic programming technique. As shown in FIG. 3, the learning network generation device 1 is realized as, for example, a computer. That is, the learning network generation device 1 includes the processor 101, the RAM 102, the HDD 103, the graphic processing unit 104, and the input interface 105.

  The entire learning network generation device 1 is controlled by the processor 101. Processor 101 may be a multiprocessor. The processor 101 is, for example, a central processing unit (CPU), a micro processing unit (MPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a graphics processing unit (GPU), or a programmable logic device (PLD). . Further, the processor 101 may be a combination of two or more elements among a CPU, an MPU, a DSP, an ASIC, and a PLD.

  The RAM 102 (Random Access Memory) is used as a main storage device of the learning network generation device 1. The RAM 102 temporarily stores at least a part of an OS (Operating System) program and an application program to be executed by the processor 101. Further, the RAM 102 stores various data necessary for processing by the processor 101.

  The HDD 103 (Hard Disk Drive) is used as an auxiliary storage device of the learning network generation device 1. The HDD 103 stores an OS program, an application program, and various data. It should be noted that another type of non-volatile storage device such as a solid state drive (SSD) can be used as the auxiliary storage device.

  A display device 104a is connected to the graphic processing unit 104. The graphic processing unit 104 displays an image on the screen of the display device 104a according to a command from the processor 101. As the display device 104a, a liquid crystal display, an organic EL (Electro Luminescence) display, or the like is used.

  An input device 105a is connected to the input interface 105. The input interface 105 transmits a signal output from the input device 105a to the processor 101. Examples of the input device 105a include a keyboard and a pointing device. Examples of the pointing device include a mouse, a touch panel, a tablet, a touch pad, and a trackball.

  With the above hardware configuration, the learning network generation device 1 can be realized.

[Functional Blocks of Learning Network Generation Apparatus 1]
FIG. 4 shows a functional block diagram of the learning network generation device 1.
The learning network generation device 1 includes a processing unit 11, a storage unit 12, an input unit 13, and an output unit 14.

  The processing unit 11 is realized by the processor 101 executing a program stored in the RAM 102 or the HDD 103. The storage unit 12 is realized by the RAM 102 and the HDD 103. The input unit 13 is realized by the input interface 105 and the input device 105a. The output unit 14 is realized by the graphic processing unit 104 and the display device 104a.

  As illustrated in FIG. 4, the processing unit 11 includes an initial individual generation unit 111, an individual code generation unit 112, a learning execution unit 113, an individual evaluation / selection unit 114, and an evolution execution unit 115 according to the processing content. Sub-blocks. As shown in FIG. 3, the storage unit 12 stores a main program 121, an individual code generation program 122, a layer name definition data 123, a layer structure definition data 124, a layer code generation rule 125, template data 126, and an individual code 127. And various programs and data including the learning data 128 are stored.

  The main program 121, when executed by the processor 101, realizes functions of the initial individual generation unit 111, the learning execution unit 113, the individual evaluation / selection unit 114, and the evolution execution unit 115. The individual code generation program 122 implements the function of the individual code generation unit 112 by being executed by the processor 101. The main program 121 and the individual code generation program 122 will be described later in detail.

  The layer name definition data 123 and the layer structure definition data 124 define a rule for extracting a layer node LN from character string data of an individual and generating a layer object.

  The layer object is obtained by applying the value of the parameter node PN of the layer node LN to the layer node LN of each layer in the learning network represented by the tree structure. It is converted into one layer object together with the parameter node PN for LN. Based on this layer object, a program code corresponding to each layer node LN is obtained. One individual is represented as an array of layer objects.

  The layer name definition data 123 defines the names of the layer nodes LN handled by the learning network generation device 1, the names and the order of the parameter nodes PN connected to each layer node LN, and the like. The layer structure definition data 124 defines a data structure of a layer object for each type of the layer node LN. That is, the type and number of the parameter nodes PN, the significance of the values of the parameter nodes PN, and the like differ for each type of the layer node LN, and the layer structure definition data 124 defines such information for each type of the layer node LN. The data structure defined in the layer structure definition data 124 includes, for example, data indicating the number of units included in the layer object, data indicating the type of the activation function (for example, identification number), data indicating a parameter, and the like.

  In the implementation in the program, the data structure according to the type of the layer node LN is described in the layer structure definition data 124 as a member variable in a class defined according to the type of the layer node LN. In addition, the class defined according to the type of the layer node LN may also define member functions such as a constructor for creating an instance of the layer object and a destructor for destroying the instance. Note that there is also a layer node LN having no parameter node PN, such as 'Flatten', and a class defining such a layer node LN having no parameter node PN does not include a definition relating to the parameter node PN.

  The layer code generation rule 125 defines rules and conditions for converting a layer object into a specific program code character string. For example, the layer code generation rule 125 defines a conversion rule that associates data indicating the type of the activation function of the instance of the layer object with a specific activation function. Further, the layer code generation rule 125 defines a rule for converting an instance of a layer object into a program code that defines processing in the layer.

  The template data 126 defines a code commonly included in all the individual codes. The template data 126 defines, for example, a part other than the learning network part of the deep learning. In the case of using a form of an external input to the generated learning network, the input layer of the learning network is common to all the generated learning networks. As described above, when there is a portion common to all the learning networks, such a common portion may be defined in the template data 126. In this way, the changeable part and the common part in the learning network can be completely separated and handled. Therefore, the amount of data to be processed in the changeable portion can be reduced. Further, the common part can be collectively modified / changed.

  The individual code 127 is an individual code of each individual generated by the individual code generation unit 112, and is a program code described in a character string. The individual code 127 is generated based on information of an individual represented as an array of layer objects, layer name definition data 123, layer structure definition data 124, layer code generation rules 125, and template data 126.

  The learning data 128 is learning data used when the individual code 127 performs deep learning. The learning data 128 is defined, for example, by a set of input data and correct answer data.

  Next, a learning network generating apparatus is used to describe how to implement the initial individual generating unit 111, the individual code generating unit 112, the learning executing unit 113, the individual evaluating / selecting unit 114, and the evolution executing unit 115, which are the components of the processing unit 11. 1 will be described together with the flow of the learning network generation process. FIG. 5 is a flowchart illustrating the procedure of the learning network generation process.

  The initial individual generation unit 111 is realized by the processor 101 executing the main program 121. The main program 121 includes, for example, a description rule of a learning network handled by the learning network generation device 1 such as a connection rule of the layer node LN and an allowable number of layers, and initial individual generation conditions such as the number of initial individuals to be generated. . The initial individual generation unit 111 generates a character string that defines the structure of the learning network according to the initial individual generation rule defined in the main program 121, and outputs the character string to the individual code generation unit 112 (Step S100).

  The individual code generation unit 112 is realized by the processor 101 executing the individual code generation program 122. The individual code generation unit 112 generates an individual code 127 of the individual based on the character string data representing the learning network of each individual. The processing performed by the individual code generation unit 112 is as follows.

  First, the individual code generation unit 112 decomposes a character string defining an individual for each layer into an array of layer objects (step S110). Specifically, in the character string defining the individual, the individual code generation unit 112 converts the character string indicating the layer node LN defined in the layer name definition data 123 to the character string immediately before the character string indicating the next layer node LN. Are recognized as a unit representing one layer object. Then, an instance of the layer object corresponding to the name of the layer node LN is generated according to the data structure of the layer object defined in the layer structure definition data 124. At this time, the value of the parameter node PN is substituted as the value of the corresponding member variable in the instance.

  In this way, the individual code generation unit 112 sequentially converts all the layer nodes LN included in the character string defining the individual into instances of the layer object, and stores them in an array arranged in the conversion order.

  Subsequently, the individual code generation unit 112 generates a program code corresponding to the learning network of the individual based on the array of the layer objects based on the layer code generation rule 125, and defines the generated program code in the template data 126. The individual code 127 of the individual in the learning network is generated and output by combining the common code and the common code (step S120). The individual code 127 is a program code described in a character string, and is recorded in the learning network generation device 12 for later learning and evaluation.

  The learning execution unit 113 is realized by the processor 101 executing the main program 121, similarly to the initial individual generation unit 111. The learning execution unit 113 executes the deep learning by applying the learning data 128 to the individual code 127 (step S130). That is, the learning execution unit 113 first sets a random value as a random value as an initial value of the weight coefficient for the input of each unit. Then, the learning input data specified in the learning data 128 is applied to the individual code 127, and the output of the individual code 127 is compared with the correct answer data corresponding to the input data specified in the learning data 128. Then, the internal parameters of the learning network described by the individual code 127 are adjusted according to the matching result. By repeating such processing, the learning network can perform deep learning. The conditions for performing deep learning (for example, the number of repetitions) are defined in the main program 121. As described above, since the initial value of the weight coefficient is a random value, the learning result may be different each time the same individual code 127 is executed. For this reason, it is preferable that a plurality of deep learnings are performed on one individual code 127 and the best learning result is adopted.

  The individual evaluation / selection unit 114 is realized by the processor 101 executing the main program 121, similarly to the initial individual generation unit 111. The individual evaluation / selection unit 114 evaluates each learned individual on which the learning execution unit 113 has performed the deep learning. As a method of evaluation, various known evaluation functions can be used.

  Subsequently, the individual evaluation / selection unit 114 performs an individual selection process based on the evaluation result. The main program 121 includes, for example, a determination condition for determining whether to leave each individual according to the evaluation result. The individual evaluation / selection unit 114 selects the individual to be left based on the evaluation result and the determination condition, and deletes the other individuals (step S140).

  The evolution execution unit 115 is realized by the processor 101 executing the main program 121, similarly to the initial individual generation unit 111. Prior to the evolution execution unit 115, the main program determines whether to execute the evolution process (step S150). Whether or not to execute the evolution process may be determined based on, for example, the number of times the evolution process has been performed from the initial individual up to now, the evaluation results of the remaining individuals, and the like. When it is determined that the evolution process is to be performed (Step S150; Yes), the evolution execution unit 115 performs the evolution process on the remaining individuals (Step S160). That is, the evolution execution unit 115 performs the crossover process and the mutation process on the individual.

  As the crossover processing, for example, crossover may be performed by a one-point method. In the one-point method, two individuals are selected as parent individuals from the remaining individuals, and as shown in FIG. 6, a layer that can be exchanged without violating the connection rule from each of the two selected parent individuals. The node LN is selected, and the learning network is disconnected at the layer node LN, and reconnected to be a child individual. Note that three or more child individuals may be generated by performing different crossover processes on the two parent individuals.

  In addition, as the mutation process, for example, a part of the remaining individual (parent individual) is randomly generated (within a range that does not violate the connection rule) by newly generating an individual in the same manner as when generating the initial individual. It is preferable to adopt a method of changing the pattern or adding a new layer randomly to the remaining individual (parent individual).

  The evolution execution unit 115 repeatedly performs the evolution process while randomly changing parent individuals until the number of child individuals defined in the main program 121 is generated. The crossover process and the mutation process by the evolution execution unit 115 may be performed on tree-structured data described by the original character string corresponding to the individual evaluated by the individual evaluation / selection unit 114. This makes it possible to easily apply the same connection rules as when generating the initial individual and the rules for generating the program code of the individual during evolution. Further, the tree-structured data described by the character string does not include the program code itself, but includes only information indicating the structure of the learning network, so that the processing load performed by the processing unit 11 can be suppressed.

  After the completion of the evolution process, the process moves to step S120. That is, it is determined that deep learning in the learning execution unit 113, evaluation and selection in the individual evaluation / selection unit 114, and evolution in the evolution execution unit 115 do not execute a further evolution process on the generated child individual. The process is repeated until a condition (for example, the number of repetitions) is satisfied. If it is determined that the evolution process is not to be performed (Step S150; No), among the remaining individuals at that time, those with excellent evaluation results are output as finally generated individuals (Step S170). The learning network generation processing ends. The number of individuals finally generated may be one or plural.

  With the configuration and procedure described above, the learning network generation device and the learning network generation program of the present embodiment can automatically generate a learning network having a structure suitable for a target.

  Although the present embodiment has been described above, the present invention is not limited to these examples. For example, in the above embodiment, the learning network is defined by a tree structure as shown in FIG. The tree structure in FIG. 2 includes one layer node LN for each layer. However, one layer may include a plurality of layer nodes LN, and a learning network having a complicated structure with branching may be defined.

  Further, in the above embodiment, the learning data 128 is defined by a set of the input data and the correct answer data. However, when performing the so-called “unsupervised learning” in the learning executing unit 113, the learning data 128 does not include the correct answer data. You may.

  Also, those in which a person skilled in the art appropriately adds, deletes, or changes the design of the above-described embodiments, or appropriately combines the features of the embodiments also have the gist of the present invention. As long as they are included in the scope of the present invention.

Reference Signs List 1 learning network generation device 11 processing unit 12 storage unit 13 input unit 14 output unit 111 initial individual generation unit 112 individual code generation unit 113 learning execution unit 114 individual evaluation / selection unit 115 evolution execution unit 121 main program 122 individual code generation program 1231 Layer name definition data 1232 Layer structure definition data 127 Code 124 Template data 125 Individual code 126 Learning data

Claims (8)

An initial individual generation unit that generates data defining an individual of the learning network;
Based on data defining the individual of the learning network generated by the initial individual generation unit, a code generation unit that generates a program code corresponding to the individual,
A learning execution unit that executes evolutionary learning by applying learning data to a program code corresponding to an individual generated by the code generation unit;
Evaluating the learned individual that has performed the evolutionary learning in the learning execution unit, and an evaluation / selection unit that determines whether to leave the individual according to the evaluation result,
A learning network generation device, comprising:   The learning network generation device according to claim 1, wherein the initial individual generation unit generates data representing a structure of the learning network. The initial individual generation unit, data representing the structure of the learning network,
A layer node corresponding to one layer in the learning network and indicating a connection relationship between at least the one layer and a previous layer;
The learning network generation device according to claim 2, wherein the learning network generation device outputs the data as a tree structure including a parameter node indicating a parameter related to a structure of a layer corresponding to the layer node. The learning network has a structure in which a plurality of layers each having one or more units are connected,
4. The learning network generation device according to claim 2, wherein the data representing the structure of the learning network includes the number of units constituting each layer, an activation function in each unit, and a connection relationship between layers. 5.   The code generator generates a program code corresponding to an individual based on data representing a structure of a learning network generated by the initial individual generator and template data of a common part independent of the individual. The learning network generation device according to claim 1. The individual selected by the evaluation / selection unit further includes an evolution execution unit that performs an evolution process to generate a next generation individual,
The learning network generation device according to any one of claims 1 to 5, wherein the code generation unit generates a program code according to the next generation individual generated by the evolution execution unit.   A learning network generation program for causing a computer to function as the learning network generation device according to any one of claims 1 to 6. On the computer,
An initial individual generation step of generating data defining an individual of the learning network;
A code generation step of generating a program code corresponding to the individual based on the data defining the individual of the learning network generated in the initial individual generation step,
A learning execution step of executing evolutionary learning by applying learning data to a program code corresponding to the individual generated in the code generation step,
Evaluating the learned individual that has performed the evolutionary learning in the learning execution step, and an evaluation / selection step of determining whether to leave the individual according to the result of the evaluation,
A learning network generation program, characterized by causing a computer to execute a learning network.

Images