JP2007087055A - Evolution calculation system and evolution calculation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To increase the degree of freedom of reproduction such as crossing processing or mutation processing, and to simultaneously optimize an individual structure and numeric value in an evolution calculation system for operating evolution calculation by genetic programming. <P>SOLUTION: This evolution calculation system is provided with a genetic table for defining the number of connection of gene, and when expression type individual structure information describing a connection relation between genes is generated, gene as the node of connection destination is specified so that the number of connection can be satisfied, and an expression type is consistently configured. Also, a parameter to be used for a function to be specified by gene is added to gene information, and the gene and the parameter are simultaneously operated. Furthermore, node classification is added to gene information, and termination and non-termination are operated, and a partial tree structure is dynamically changed. A matrix type individual structure is adopted, and the operation of the crossing or mutation of a block is carried out by reproduction processing. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an evolutionary computation system that performs evolutionary computation by genetic programming, and relates to a technique for improving the degree of freedom of reproduction in genetic programming and further simultaneously optimizing the individual structure and numerical values.

  Evolutionary computation is an optimization (search) algorithm inspired by biological evolution. It can be said to be a kind of multi-point search method that operates on data based on evolutionary ideas and searches the solution space. Typical methods include genetic algorithms and genetic programming.

  The genetic algorithm generally defines an individual as a one-dimensional string (chromosome) of bits indicating 0 or 1. Then, as reproduction, crossover processing for exchanging chromosomal sites between two individuals or mutation processing for randomly changing part of the chromosomes is performed by one-point crossover, multipoint crossover, uniform crossover, or the like.

  Genetic programming is an evolution of a genetic algorithm, and represents an individual structurally like a graph structure or a tree structure. For example, the tree structure is expressed by Lisp's S expression described as (A (B) (C (D))). Thus, since it is described as a one-dimensional character string, a new expression can be generated by crossover processing. However, there is a possibility that a lethal gene that is not actually a structure is included. There is a problem that the character string cannot be rewritten at random in the mutation process. In other words, genetic programming has a low level of voluntary reproductivity, and individual diversity cannot be ensured. If this point is improved, it is possible to improve the search efficiency related to the structurally expressed individual.

Further, the conventional evolutionary calculation method has a problem that the structure and numerical values cannot be optimized simultaneously. Since the genetic algorithm uses a symbol string composed of bits as chromosomes, it is suitable for numerical optimization, but is not suitable for structural optimization. On the other hand, genetic programming has made it possible to solve complex problems by structural representation, but numerical optimization such as parameter setting for nodes has not been achieved.
Hiroshi Sato, Isao Ono, Shigenobu Kobayashi, “Proposal and Evaluation of Generation Alternation Model in Genetic Algorithm”, Journal of Artificial Intelligence, Journal of Artificial Intelligence 1997, vol. 12, no. 5, p. 734-744 Masashi Iba, Taisuke Sato, “Genetic Programming Based on System Identification Approach”, Journal of Artificial Intelligence Society, Artificial Intelligence Society, 1995, vol. 10, no. 4, p. 100-110 Shinya Aoki, Tomoharu Nagao, “Automatic construction method of tree-structured image conversion ACTIT”, The Institute of Image Information and Television Engineers, The Institute of Image Information and Television Engineers, 1999, vol. 53, no. 6, p. 888-894

  SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to solve the above-mentioned problems, to improve the reproductive arbitraryness in genetic programming, and to simultaneously optimize the structure of an individual and the numerical values used in nodes.

The evolutionary computation system according to the present invention is:
An evolutionary computation system that performs evolutionary computation using genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and has the following elements: (1) Individual population related to the same generation An individual population storage unit (2) that stores genotype individual structure information of each individual that is an element of the gene table (2) A gene table that defines the number of data input by the function specified by the gene as a connection number ( 3) The genes included in the genotype individual structure information are sequentially read, the number of linked genes is obtained from the gene table, the genes to be linked nodes are specified so as to satisfy the linked number, and the linkage relationship between the genes is determined. Generate the phenotypic individual structure information to be described, input the learning input data, and connect the gene linkage specified by the phenotype individual structure information. In accordance with, an fitness calculation unit that calculates the fitness of each individual by calculating the individual computation output data by sequentially executing the calculation by the function specified by each gene and comparing the individual computation output data with the target data for learning ( 4) An individual selection unit for selecting an individual from an individual population based on fitness (5) Crossover processing for exchanging a part of information among a plurality of individuals for the genotype individual structure information of the selected individual, or A reproductive unit that performs mutation processing that changes some information, obtains genotype individual structure information of each individual that is an element of the next generation individual population, and stores it in the individual population storage unit.

Further, the genetic information of the genotype individual structure information further includes a parameter added to the gene,
The fitness calculation unit is characterized in that, when performing an operation based on a function specified by a gene, the operation is performed using a parameter added to the gene.

The gene information of the genotype individual structure information further includes a node type indicating whether the gene is a terminal node or a non-terminal node when the gene is a node constituting a connection relationship,
When generating the phenotype individual structure information describing the linkage relationship between genes, the fitness calculation unit sets the number of linked genes to 0 when the node type of the gene indicates the end node. It is characterized by.

In addition, the genetic information of the genotype individual structure information includes at least a gene, a parameter added to the gene, and whether the gene is a terminal node or a non-terminal node when the gene is a node constituting a connection relationship. Is a record including an item of a node type, and the genotype individual structure information is matrix type individual structure information constituting a matrix made up of records and items,
The genital part is characterized by performing a crossover process for exchanging blocks across a plurality of records as a part of a matrix or a mutation process for changing a block across a plurality of records as a part of a matrix.

In addition, the genetic information of the genotype individual structure information is a terminal node when at least a bit configuration indicating a gene, a bit configuration indicating a parameter added to the gene, and a node that forms a connection relationship with the gene. It is a record including a bit configuration indicating a node type indicating whether it is a non-terminal node, genotype individual structure information is matrix type individual structure information that constitutes a matrix consisting of records and bits,
The genital part is characterized by performing a crossover process for exchanging blocks across a plurality of records as a part of a matrix or a mutation process for changing a block across a plurality of records as a part of a matrix.

The evolution calculation method according to the present invention is:
For each individual that is an element of an individual group related to the same generation, an individual group storage unit that stores genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and an associated gene and specified by the gene An evolution calculation method using an evolution calculation system that has a gene table that defines the number of data input by the function to be input as the number of connections and performs evolution calculation using genotype individual structure information, and has the following elements (1) sequentially reading the genes included in the genotype individual structure information, obtaining the number of linked genes from the gene table, specifying the genes that are the nodes to be linked to satisfy the linked number, Generate phenotypic individual structure information describing the linkage relationship between genes, input learning data, and use phenotypic individual structure information In accordance with the specified gene linkage relationship, the calculation by the function specified by each gene is sequentially executed to obtain individual calculation output data, and the individual calculation output data is compared with the learning target data to determine the fitness for each individual. The fitness calculation step for calculating (2) the individual selection step for selecting an individual from the individual population based on the fitness (3) part of information among a plurality of individuals for the genotype individual structure information of the selected individual A reproductive process in which crossover processing for exchanging or mutation processing that changes some information is performed to obtain genotype individual structure information of each individual that becomes an element of the next generation individual population and to be stored in the individual population storage unit.

A computer-readable recording medium according to the present invention includes:
For each individual that is an element of an individual group related to the same generation, an individual group storage unit that stores genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and an associated gene and specified by the gene The computer has a gene table that defines the number of data input by the function to be input as the number of connections, and uses the genotype individual structure information to perform evolutionary calculations. It is a computer-readable recording medium in which a program is recorded. (1) Sequentially reads genes included in genotype individual structure information, obtains the number of linked genes from the gene table, and satisfies the number of linkages. The phenotypic individual structure information that specifies the gene that is the node to be linked to and describes the linkage relationship between the genes Generate and input learning data, and execute individual functions according to the linkage relationship between the genes specified by the phenotype individual structure information to obtain individual calculation output data. The fitness calculation process for calculating the fitness for each individual by comparing the output data with the learning target data (2) The individual selection process for selecting the individual from the individual population based on the fitness (3) The selected individual For genotype individual structure information, perform crossover processing to exchange some information between multiple individuals, or mutation processing to change some information, and each individual that will be an element of the next generation individual population Reproductive processing for obtaining genotype individual structure information and storing it in the individual population storage unit.

The program according to the present invention is:
For each individual that is an element of an individual group related to the same generation, an individual group storage unit that stores genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and an associated gene and specified by the gene For a computer that is an evolutionary computation system that performs evolutionary computation using genotype individual structure information, and has a gene table that defines the number of data input by the function to be input as the number of connections. (1) The genes included in the genotype individual structure information are sequentially read, the number of linked genes is obtained from the gene table, and the genes that are the nodes to be linked to satisfy the number of linkages are obtained. Generate phenotypic individual structure information that identifies and describes the linkage relationship between genes, inputs input data for learning, and phenotypes According to the linkage relationship of the genes specified by the body structure information, the calculation by the function specified by each gene is sequentially executed to obtain individual calculation output data, and the individual calculation output data is compared with the learning target data for each individual. (2) Individual selection procedure for selecting an individual from an individual population based on the fitness (3) For the genotype individual structure information of the selected individual among a plurality of individuals Perform crossover processing to exchange some information or mutation processing to change some information, obtain genotype individual structure information of each individual that will be an element of the next generation individual population, and store it in the individual population storage unit Reproductive procedure to let.

  According to the present invention, when generating the phenotypic individual structure information describing the connection relationship between genes, the gene that is the node of the connection destination is specified so as to satisfy the number of connections. There is no contradiction. Therefore, the reproductivity is improved, and individual diversity is exhibited. In particular, it is superior to conventional genetic programming in that the middle layer part structure (subtree) can be exchanged.

  Since the parameters used for the function are added to the gene information, the structure and the numerical value can be simultaneously operated in the reproduction, and both can be optimized simultaneously.

  By adding the node type to the gene information and manipulating the terminal and non-terminal, the structural parts of the gene can be dynamically changed. It is expected that this part will serve as a functional module, and that structural operations for high level functions will be performed by reproductive processing.

  Employing a matrix-type individual structure and performing block operations in reproductive processing, it is possible to bring about changes that have not been made in the past for phenotypes.

Embodiment 1 FIG.
FIG. 1 is a diagram showing a configuration of an evolutionary computing system. The evolution calculation system includes an initial individual input unit 101, an individual group storage unit 102, an fitness calculation unit 103, a gene table 104, an fitness storage unit 105, an individual selection unit 106, a selected individual storage unit 107, a reproduction unit 108, and learning. An individual output unit 109 is provided. The individual group storage unit 102 is configured to store genotype individual structure information of each individual that is an element of an individual group related to the same generation.

  The configuration of the gene table 104 will be described. FIG. 2 is a diagram illustrating an example of a gene table. This table stores the genes and the number of linked genes in association with each other and defines the relationship between them. A gene is an identifier that identifies a function that performs an operation. The number of connections indicates the number of data input to the function specified by the gene. For example, when performing image processing as a calculation, an image processing filter is specified by gene. In addition, examples of “1-input 1-output filter” in which the number of image data to be input is one and the number of image data to be output is one include an average value filter, a maximum value filter, a minimum value filter, and an inversion filter. These are defined as 1 connected. Examples of the “two-input one-output filter” in which the number of input image data is two and the number of image data to be output is one include a logical sum filter, a logical product filter, a discrimination filter, and the like. It is defined as

  In this example, the 0-linked gene is a terminating gene used as a terminal node, and the 1- or 2-linked gene is a non-terminal gene used as a non-terminal node. A terminal node is a node that directly inputs learning data that is an initial input value, and a non-terminal node is a node that inputs output data of another node. The non-terminal node is also used as a root node that outputs individual computation output data that is a final output value. In this example, even when the same calculation is performed, the terminator and the terminator are distinguished as genes. That is, the gene when the average value filter mentioned as an example of the above-mentioned “1-input 1-output filter” is used at the end and the gene when used at the non-end are set separately. In a gene using the “one-input one-output filter” as a terminal, learning input data that is an initial input value is input to the filter. For genes that use the “two-input one-output filter” at the end, the learning input data that is the initial input value is used as both input data, and the filter is executed. Alternatively, learning input data is used as one input data, and other input data is predetermined data (image data having a maximum value for all pixels, image data having a minimum value for all pixels, image data having an average value for all pixels, etc. ) To execute the filter.

  Hereinafter, the learning process by the evolution calculation algorithm by the evolution calculation system will be described. FIG. 3 is a diagram showing a flow of learning processing by the evolutionary calculation algorithm. First, an initial individual input process (S301) by the initial individual input unit 101 is performed. In this process, an initial individual group is input and stored in the individual group storage unit 102. That is, individual structure information (hereinafter referred to as an individual structure) for specifying the structure of an individual is input for the number of elements constituting the group, and individual identification information (hereinafter referred to as an individual ID) is associated with them. And stored in the individual group storage unit 102.

  Here, the genotype individual structure input in the present embodiment will be described. FIG. 4 is a diagram showing the genotype individual structure of individual i. In ascending order of the node numbers, the genes constituting the nodes are stored as columns. In this form, the genetic information is composed only of genes, but the genotype individual structure has a structure in which genetic information including at least genes is continuous. In the form described later, other information is added to the gene information.

  Next, fitness calculation processing (S302) by the fitness calculation unit 103 is performed. In this process, the fitness of each individual stored in the individual group storage unit 102 is calculated. The fitness is the learning data that is the initial input value, and the individual calculation output data that is the result of calculation according to the calculation procedure specified by the individual structure is the final target data for learning. Information indicating the degree of approximation. The fitness is stored in the fitness storage unit 105 in association with the individual ID. Details of the fitness calculation process (S302) will be described later.

  Next, an individual selection process (S303) by the individual selection unit 106 is performed. In this process, based on the fitness of each individual, a predetermined number of selected individuals are selected for reproductive processing. For example, a selected number of individuals are selected in descending order of fitness. Alternatively, the number of selections is divided into the number of highly adaptive individuals selected in the high fitness layer and the number of low adaptive individuals selected in the low fitness layer, and the individuals with the high adaptive individual selection number are selected in descending order of fitness. Further, the fitness may be stratified so that individuals with a low adaptive individual selection number are selected in order of lower fitness, and a predetermined number of each may be selected from each layer. In addition, a roulette selection method that selects individuals with a probability proportional to the fitness, a rank selection method that determines the number of selections according to the fitness ranking, and a random number of individuals selected as the tournament size from the population. There is also a tournament selection method in which an individual with the highest fitness among them is selected as a tournament winner. In this way, the ID of the selected individual group is stored in the selected individual storage unit 107.

  Next, a reproduction process (S304) by the reproduction unit 108 is performed. In this process, a configuration related to generational change is applied to an individual extracted from the selected individual group. In this example, a crossover process for exchanging a part of a gene string between two individuals, a mutation process for changing a part of the gene string, and a life extension process for maintaining the current solid structure are performed. Through these processes, a next-generation individual group consisting of individuals of a predetermined number of next-generation individuals is generated, and an individual structure is stored in the individual group storage unit 102 in association with each individual ID. Details of this reproductive process (S304) will be described later.

  Then, these generation change processes (S302 to S304) are repeated, and when the generation change end determination unit (not shown) determines that the generation change end determination process (S305) continues, the generation change process loops and ends. If it is determined, the generation change process loop is terminated and the process proceeds to post-processing. The end of the generation change is determined by the number of generations, for example. The number of generation change processes is counted, and when the number of generations is less than the predetermined number of generations, it is determined to be continued, and when the number of generations is reached, it is determined to end. Alternatively, it can be determined that the fitness is complete when the fitness satisfies the termination condition. For example, the target fitness and the target number of individuals are set as end conditions, the number of individuals that have reached the target fitness is counted, and if the number does not reach the target number of individuals, it is determined to continue and the target number of individuals has been reached. It is determined that the process ends. Alternatively, set a time limit for generation change processing, measure the processing time between the loops of generation change processing, determine if the processing time does not reach the time limit, and decide to end when the time limit is reached To do.

  As post-processing, fitness calculation processing (S306) by the fitness calculation unit 103 is performed on the individual population of the last generation as described above, and learning individual output processing (S307) by the learning individual output unit 109 is performed. In the learning individual output process, the individual structure of each individual included in the individual population of the last generation and the fitness of the individual are associated and output. That is, for each individual ID stored in the individual group storage unit 102, the individual structure corresponding to the ID is extracted, the fitness associated with the ID is extracted from the fitness storage unit 105, and the extracted individual The structure and fitness are associated with each other and output as learning individual information. At that time, it is also effective to sort the individual information group using the fitness as a key and output the sorting result.

  Hereinafter, the fitness calculation process (S302) and the reproductive process (S304) will be described in detail.

  First, the fitness calculation process (S302) shown in FIG. 3 will be described. FIG. 5 is a diagram illustrating a configuration of the fitness calculation unit. The fitness calculation unit 103 includes a type conversion unit 501, an individual calculation unit 502, and a data comparison unit 503.

  FIG. 6 is a diagram illustrating a flow of fitness calculation processing. The type conversion process (S602), the individual calculation process (S603), and the data comparison process (S604) are performed on all the individuals stored in the individual group storage unit 102. In the type conversion process (S602) by the type conversion unit 501, the genotype individual structure is converted into phenotype individual structure information (hereinafter referred to as phenotype individual structure). The phenotypic individual structure has a structure that describes the linkage relationship between genes. In the individual calculation process (S603) by the individual calculation unit 502, input data for learning which is an initial input value is input, and a series of calculations are performed according to the calculation procedure specified by the phenotype individual structure to obtain individual calculation output data. . For example, when performing image processing, the original image data to be processed is input, the image data is processed by processing with the image processing filter sequentially connected, and the processed image data after the image processing is processed as an individual image. Output as operation output data. In the data comparison processing (S604) by the data comparison unit 503, the individual calculation output data and the learning target data are compared, and the fitness is calculated according to a predetermined standard. Note that the learning input data and the learning target data are input in advance via a learning data input unit (not shown) and stored in a learning data storage unit (not shown).

  Hereinafter, the type conversion process (S602), the individual calculation process (S603), and the data comparison process (S604) will be described in detail.

  First, the type conversion process (S602) shown in FIG. 6 will be described. FIG. 7 is a diagram showing a flow of type conversion processing. In this process, the genotype individual structure shown in FIG. 4 is converted into a phenotype individual structure. FIG. 8 is a diagram showing a process of generating the phenotype individual structure of the individual i. The phenotype individual structure has a record for each node, and is configured to store a gene, the number of connections, the first connection destination node No, and the second connection destination node No in association with the node No.

  As shown in FIG. 6, the following processing is sequentially repeated for each node (S701). If the node is No = 1 (root) (S702), S703 to S705 are omitted, and the process proceeds to S706. Genes are read in order from the genotype individual structure (S706), the number of linked genes is obtained from the gene table (S707), and the number of genes and linkages are written in the record of the node of the phenotype individual structure (S708). . Further, reservation codes are written in the connection destination nodes corresponding to the number of connections of the record (S709). In the example of the individual A, the number of connections “2” of the gene “E” is acquired from the gene table, and each is set in the first record of the phenotype individual structure. Since the number of connections is “2”, reservation codes are set in the first connection destination node No and the second connection destination node No. The reservation code indicates that it is necessary to set a gene in the node to be linked.

  If there is a next gene in the genotype individual structure (S710), the processing from S703 onward is repeated for the next node. First, the position to connect is specified. For this purpose, a reservation code to be linked is specified from the linked nodes in each record of the phenotype individual structure (S703). There are various methods for specifying which reservation code when there are a plurality of reservation codes.

  In this example, priority is given to the item of the 1st connection destination node No, and when the reservation code exists in the item of the 1st connection destination node No of any record, the reservation code is made into connection object. If there is no reservation code in the item of the first connection destination node No of any record, select the reservation code of the lowest record among the reservation codes in the second connection destination node No. Consolidate. The reservation code specifying method according to this example is a depth-first specifying method.

  In addition, there is a method of specifying width priority. In the width priority specifying method, the reservation code of the highest record is selected. When there are two reservation codes in the highest record, the first connection destination node No is selected.

  If there is no reservation code, the data input source has been determined for all nodes, and the process is terminated (S704).

  When any reservation code is specified as a connection target, the specified reservation code is rewritten to No of the relevant node (S705). As a result, the node is linked to the node of the record having the reservation code. In the example of FIG. 8, the second node is connected to the first node. Similarly, the third node is connected to the second node, and the fourth node is connected to the first node. This indicates that the first node inputs the output data of the second node and the output data of the fourth node, and the second node inputs the output data of the third node. Since the third node is not connected, it indicates that learning input data that is an initial input value is directly input.

  When all the genes of the genotype individual structure have been processed, the process ends as incomplete (S710). In the case of incomplete, for example, recovery processing is performed such that the individual is excluded from the subsequent processing target, or a predetermined termination node is assigned to the reservation code and forcibly completed.

  In this manner, the individual phenotype individual structure shown in FIG. 9 is obtained. In addition, FIG. FIG. 10 is a diagram showing an image of the phenotype of individual i.

  Next, the individual calculation process (S603) shown in FIG. 6 will be described. FIG. 11 is a diagram illustrating a flow of the individual calculation process. First, a node to be calculated is searched (S1101). Specifically, any node that can be calculated is selected. The condition of a node that can be calculated is that data to be input to the node is fixed. Since the data to be input is determined for the node having the connection number of 0, it can be calculated immediately. A node having the number of connections of 1 or more can be calculated when output data from all nodes to which the connection is made is obtained. As will be described later, the calculation result data that is the output of each node is stored in a calculation result storage unit (not shown) in the individual calculation unit in association with the ID of the node. When all the operation result data to be stored is stored in the operation result storage unit, it is determined that the node can be calculated. When there are a plurality of nodes that can be calculated, which node is subject to calculation is not particularly limited, but in this example, the lower nodes are preferentially selected.

  Next, the function processing unit corresponding to the gene of the node to be calculated is specified (S1102). Since a gene is an identifier for identifying a function to perform an operation, a function processing unit can be determined according to the definition. In the case of image processing, the image processing filter corresponds to the function processing unit. Next, data to be input to the node is specified (S1103). In the case of a terminal node with a connection number of 0, learning input data is used as input data to the node. If the end node has specifications that also use predetermined data (image data with the maximum value for all pixels, image data with the minimum value for all pixels, image data with the average value for all pixels, etc.) To do. In the case of a non-terminal node having a connection number of 1 or more, operation result data associated with each connection destination node No and stored in the operation result storage unit is input data to the node.

  Then, the specified data is input to perform calculation by the function processing unit (S1104), and the calculation result data is stored in the calculation result storage unit in association with the node number of the node (S1105).

  The above processing is performed for all nodes (S1106), and the calculation result data of the root (node No = 1) is used as individual calculation output data (S1107).

  In the data comparison process (S604) shown in FIG. 6, the individual calculation output data is compared with the learning target data, and the fitness is calculated according to a predetermined standard. The fitness is an approximate degree determined according to data characteristics. In general, the difference between the data is obtained, and the fitness is increased as the difference is smaller. In the case of image processing, for example, the luminance difference for each pixel is calculated, and the smaller the sum of the differences, the greater the fitness is, or the luminance difference for each pixel is calculated and the difference exceeds the reference value. There is a method of counting the number of pixels, and determining that the fitness is higher as the number of pixels having a larger luminance difference is smaller.

  Next, the reproduction process (S304) shown in FIG. 3 will be described. FIG. 12 is a diagram showing a flow of reproductive processing. First, an individual number determination unit (not shown) determines the number of individuals obtained by crossover processing, the number of individuals obtained by mutation processing, and the number of individuals obtained by life extension processing (S1201). The total number of these individuals is the total number of next-generation individuals. There are a method of using a predetermined constant and a method of determining each individual number based on a random number or the like.

  A crossover process (S1202) for exchanging a part of a gene sequence between two individuals by a crossover part (not shown), and a mutation process (S1202) for changing a part of a gene by a mutation part (not shown) In step S1203), the life extension unit (not shown) performs life extension processing (S1204) for maintaining the current individual structure. These processes are independent, and the order of the processes is not necessarily limited. When a next generation individual is obtained by these processes, an individual group update unit (not shown) performs an individual group update process (S1205). In this process, the individual ID and the genotype individual structure relating to the current generation individual are deleted, and the genotype individual structure is stored in the individual group storage unit 102 in association with the individual ID for each next generation individual. That is, the current generation individual population is updated to the next generation individual population.

  The crossover process (S1202) will be described. FIG. 13 is a diagram showing the flow of the crossover process. The process of obtaining new individuals by crossover is repeated until the predetermined number of individuals is reached (S1301). First, two individuals to be crossed are identified (S1302). Two individuals are selected from the selected individuals stored in the selected individual storage unit 107. Generally, the selection is made irregularly using random numbers or the like. There are methods that allow the same individual to be selected repeatedly, and methods that do not repeatedly select. Next, the number of crossover points is specified (S1303). When there is one crossing point, all the columns after the crossing point are exchanged. When there are two crossing points, the partial rows between the crossing points are exchanged. When there are three or more crossing points, a plurality of rows are exchanged. The number of crossover points may be a predetermined fixed number, or the irregular number may be obtained each time using a random number or the like. The number of intersections corresponding to the number is selected (S1304). The position of the crossover point is determined irregularly using random numbers or the like within the range of the maximum number of genes. There are two methods for determining the position of the crossing point: one is common to two individuals, and the other is independent. Then, the gene sequence between the crossing points is exchanged (S1305). If the number of crossover points is odd, the last sequence is exchanged between the first and second crossover points, the third and fourth crossover points, and the next crossover point in order. Exchange from the crossover point to the final gene. In general, two crossed individuals are left in the next generation, but there is also a method of leaving only one of them. In this way, an individual obtained by crossover is specified, and the genotype individual structure is temporarily stored in the reproductive part (S1306). These processes are repeated, and the process ends when a predetermined number of individuals is obtained (S1307).

  In this example, some information is exchanged between two individuals, but information may be exchanged between three or more individuals.

  The number of elements in the row to be crossed is one or more, and the elements alone, that is, one gene may be exchanged.

  An example is shown in which the above-mentioned individual A is crossed with the individual B. FIG. 14 is a diagram showing a genotype individual structure of an individual B, and FIG. 15 is a diagram showing an image of an individual B phenotype. In this example, there are two crossover points, and the gene sequence between the fourth gene and the eighth gene is exchanged. That is, the gene strings in the frame 401 in FIG. 4 and the frame 1401 in FIG. 14 are exchanged. In the phenotype image, the part structures (subtrees) in the frame 1001 in FIG. 10 and the frame in FIG. 15 are exchanged.

  As a result, the individual i changes like an individual c. FIG. 16 is a diagram showing the genotype individual structure of the individual c, and FIG. 17 is a diagram showing an image of the phenotype of the individual c. As shown in FIG. 17, while maintaining the upper part structure (the structure from node 1 to node 3) and the lower part structure (the structure of node 9 and node 10, the structure from node 11 to node 14), The parts structure (from node 4 to node 8) has been successfully replaced.

  Next, the above-described mutation process (S1203) will be described. FIG. 18 is a diagram showing a flow of mutation processing. The process of obtaining new individuals by mutation is repeated until the predetermined number of individuals is reached (S1801). First, an individual to be mutated is specified (S1802). One individual is selected from the selected individuals stored in the selected individual storage unit 107. Generally, the selection is made irregularly using random numbers or the like. There are methods that allow the same individual to be selected repeatedly, and methods that do not repeatedly select. Next, the number of mutation ranges is specified (S1803). The number of mutation ranges may be a predetermined fixed number, or an irregular number may be obtained each time using a random number or the like. The number of mutation ranges is selected (S1804). The size of the range is one or more, and may be a predetermined fixed number, or the random number may be obtained each time using a random number or the like. The position of the range is determined irregularly using random numbers or the like. Next, an alternative gene string to be replaced is specified (S1805). An alternative gene sequence uses an irregularly selected gene example. Generally, it is determined each time using a random number or the like. There are a method of making the size of the alternative gene sequence the same as the mutation range and an arbitrary method. Then, the mutation range of the genotype individual structure is replaced with an alternative gene string (S1806). In this way, an individual obtained by mutation is specified, and the genotype individual structure is temporarily stored inside the reproductive part. These processes are repeated, and the process is terminated when a predetermined number of individuals is obtained (S1807).

  The number of elements in the row to be mutated is one or more, and there are cases where the element alone, ie, one gene is replaced.

  Next, the life extension process (S1204) will be described. FIG. 19 is a diagram illustrating a flow of life extension processing. Until the predetermined number of individuals is obtained, the process of obtaining new individuals by extending the life is repeated (S1901). One individual is selected from the individuals stored in the selected individual storage unit 107. Generally, the selection is made irregularly using random numbers or the like. There are methods that allow the same individual to be selected repeatedly, and methods that do not repeatedly select. As an individual to prolong the life, the genotype individual structure is temporarily stored inside the reproductive part. This process is repeated, and ends when a predetermined number of individuals is obtained (S1903).

  In the individual group update process (S1205), a new individual ID is added to the genotype individual information temporarily stored in the reproductive part by the crossover process, the mutation process, and the life extension process, and the individual ID and genotype are added. The individual information is associated and stored in the individual group storage unit 102, and updated to the next generation individual group.

  In the present embodiment, as shown in FIGS. 10, 15, and 17, a feature appears in that the middle part structure (partial tree) can be replaced. In the figure, crossover is described as an example, but the same is true in the case where mutation can replace the middle part structure.

  Here, the operation of the middle part structure by the conventional method of describing the individual of phenotype will be described. Conventionally, Lisp's S-formula has been used as a method of describing individuals of phenotype. FIG. 20 and FIG. 21 show examples in which the above-mentioned individual A and individual B are described by the Lisp S-expression, respectively. In this formula, the lower hierarchy is listed in parentheses to represent multiple hierarchies. The lower part structure (independent subtree) is, for example, “(E (A) (C (A)))” corresponding to the eleventh node to the fourteenth node of the individual “a”. Shown as a column between parentheses. In this way, even if a column enclosed in parentheses is exchanged for a column enclosed in other parentheses, there is no inconsistency in the correspondence between parentheses. Therefore, it is possible to cross the lower part structures (independent subtrees) or to mutate the lower part structures. However, if the middle part structure is arbitrarily rewritten, the correspondence between the parentheses is inconsistent, so that the middle part structures are not exchanged or the middle part structure is not mutated. There may also be structural inconsistencies.

  Hereinafter, a structural contradiction caused by rewriting the middle part structure by the Lisp S-formula will be described. As the portion corresponding to the column “(F (B) (D (F (F)” of the fourth node to the eighth node of the individual A shown in FIG. 20, “(C (E (C ( E ”is selected and replaced to obtain the formula shown in FIG. 22. At this time, the portion of the individual B is selected so that the number of open parentheses is four. Therefore, there is no contradiction in the number of parentheses even if they are interchanged, but according to the equation shown in Fig. 22, the "E" gene at the seventh node is assigned to "(A)" and "(D (B)) ”and“ (E (A) (C (A))) ”occur, and the number of linkages for the gene“ E ”, which originally has two linkages, is over, and structurally In this way, in Lisp's S formula, it is not possible to arbitrarily replace the substring corresponding to the middle part structure without contradiction. A capability. For reference, the Lisp S-expression of individual wafer obtained in this embodiment, shown in FIG 23.

Embodiment 2. FIG.
In the present embodiment, a mode is described in which a parameter is added to a gene, and a gene and a parameter are simultaneously optimized using a genotype individual structure composed of a set of genes and parameter pairs (corresponding to gene information). . The parameter is used for a function that performs an operation specified by a gene. For example, the parameter is used as a coefficient indicating X to the power of n, and X, X square, and X cube are distinguished. Alternatively, it is used to identify the type of trigonometric function, and sin (X), cos (X), and tan (X) can be distinguished. That is, the contents of the calculation can be changed according to the parameters.

  FIG. 24 is a diagram showing the genotype individual structure of the individual to which the parameter is added.

  The configuration of the evolution calculation system and the learning processing flow by the evolution calculation algorithm in the present embodiment are the same as those in FIGS. 1 and 3 shown in the first embodiment.

  In the present embodiment, the type conversion processing (S602 in FIG. 6) by the type conversion unit 501 that is a component of the fitness calculation unit 103 shown in FIG. 5 is different from the above-described embodiment. FIG. 25 is a diagram illustrating a changed part of the type conversion process. Between S705 and S709 in FIG. 7, the following processing is performed instead of S706 and S708. A set of genes and parameters of the node is read from the genotype individual structure (S2501). Then, the number of linked genes is obtained from the gene table (S707), and the genes, parameters, and linked numbers are written in the record of the node of the phenotype individual structure (S2502).

  As a result, as shown in FIG. 26, the parameter is also stored in association with the node No in each record, and the phenotype individual structure of the individual A including the parameter is obtained. FIG. 27 is a diagram showing an image of an individual phenotype to which parameters are added. In this example, the parameter range is 1, 2, or 3, but it can be set according to the function. A continuous value in a predetermined range may be used.

  Further, the individual calculation processing (S603 in FIG. 6) by the individual calculation unit 502 which is a component of the fitness calculation unit 103 shown in FIG. 5 is different from the above-described embodiment. FIG. 28 is a diagram showing a changed part of the individual calculation process. Between S1103 and S1105 in FIG. 11, instead of S1104, the specified data is input, and the calculation is performed by the function processing unit using the parameters (S2801).

  Further, the crossover process (S1202) and the mutation process (S1203) included in the reproductive process shown in FIG. 12 are different from the above-described embodiment.

  FIG. 29 is a diagram showing a changed part of the crossover process. Between S1304 and S1306 in FIG. 13, in place of S1305, the sequence of gene / parameter pairs between crossover points is exchanged (S2901).

  31 is replaced by the set of columns shown in the frame of the individual genotype individual structure shown in FIG. 30 by replacing the set of rows shown in the frame of the individual genotype individual structure shown in FIG. The genotype individual structure of the individual C shown is obtained. FIG. 32 is a diagram showing an image of the phenotype of an individual C to which a parameter is added.

  FIG. 33 is a diagram showing a changed part of the mutation process. The following processing is performed between S1804 and S1807 in FIG. 18 instead of S1805 and S1806. A replacement gene / substitution parameter pair column is specified (S3301), and the mutation range is rewritten to a replacement gene / substitution parameter pair column (S3302). The substitute parameter is generated irregularly based on a random number or the like.

  In the present embodiment, genes and parameters can be manipulated simultaneously in reproductive processing, and both can be optimized simultaneously.

Embodiment 3 FIG.
In the second embodiment, the gene / parameter group sequence is the target of the reproductive process, but the gene-only sequence or the parameter-only sequence can also be combined as the reproductive process target.

  In that case, the crossover process (S1202) and the mutation process (S1203) included in the reproductive process shown in FIG. 12 are different from the above-described embodiment.

  FIG. 34 is a diagram showing a changed part of the crossover process. The following processing is performed between S1304 and S1306 in FIG. 13 instead of S1305. As a column to be crossed, a gene / parameter group column, a gene column, and a parameter column are determined (S3401). Based on a random number or the like, the column is randomly determined. Then, according to the difference, the column to be crossed between the crossing points is exchanged (S3402).

  FIG. 35 is a diagram showing a changed part of the mutation process. The following processing is performed between S1804 and S1807 in FIG. 18 instead of S1805 and S1806. In the same manner as described above, a gene / parameter group column, a gene column, and a parameter column are determined as columns to be mutated (S3501). Next, an alternative gene / alternative parameter pair column, an alternative gene column, or an alternative parameter column to be replaced according to the column is identified (S3502), and the mutation range column is determined according to the column. Is replaced with an alternative column (S3503).

Embodiment 4 FIG.
In the above-described embodiment, the termination gene used for the terminal node is defined as the number of linkages of 0, and the non-terminal gene used for the non-terminal node is defined as the number of linkages of 1 or more. In the form, a gene for identifying a specific function is used for both a terminal node and a non-terminal node without distinguishing the two in the gene table. Further, the node type is stored in association with the gene in the genotype individual structure, and the type conversion process (S601 in FIG. 6) is performed according to the node type.

  FIG. 36 is a diagram illustrating an example of a gene table. The genes of the present embodiment are both terminal and non-terminal, and the number of input data used in the function specified by the gene is defined as the number of connections.

  Next, the genotype individual structure in the present embodiment will be described. FIG. 37 is a diagram showing the genotype individual structure of an individual e. As shown in the figure, it is configured as a sequence of node types and gene pairs (corresponding to gene information). The node type designates one of two types, terminal (t) and non-terminal (n).

  The configuration of the evolution calculation system and the learning processing flow by the evolution calculation algorithm in the present embodiment are the same as those in FIGS. 1 and 3 shown in the first embodiment.

  In the present embodiment, the type conversion processing (S602 in FIG. 6) by the type conversion unit 501 that is a component of the fitness calculation unit 103 shown in FIG. 5 is different from the above-described embodiment. FIG. 38 is a diagram illustrating a changed part of the type conversion process. Between S706 and S709 in FIG. 7, the following processing is performed instead of S707 and S708. The node type of the gene is acquired from the genotype individual structure (S3801). If the node type (S3802) is non-terminal (n), the number of linked genes is acquired from the gene table (S3803) Use the number of connections. On the other hand, when the node type (S3802) is the terminal (t), the number of connections is set to 0 (S3804), and the subsequent processing is performed.

  FIG. 39 shows a process of generating a phenotype individual structure of an individual. In the first node, since the node type is non-terminal (n), the number of connections is set to “2” according to the definition of the gene table, and reservation codes are set in the first connection destination node No and the second connection destination node No. On the other hand, since the node type is the terminal (t) in the third node, the definition of the gene table is not followed, the number of connections is set to “0”, and the reservation code is not set to the connection destination node No. . In this way, the individual phenotype individual structure shown in FIG. 40 is obtained. FIG. 41 is a diagram showing an image of the phenotype of the individual D.

  The crossover process (S1202) and the mutation process (S1203) included in the reproductive process shown in FIG. 12 are different from the above-described embodiment.

  FIG. 42 is a diagram showing a changed part of the crossover process. The following processing is performed between S1304 and S1306 in FIG. 13 instead of S1305. As a column to be crossed, a node type / gene pair column, a node type column, and a gene column are determined (S4201). The columns are determined irregularly based on random numbers or the like. Then, the row between the crossing points is exchanged (S4202).

  FIG. 43 is a diagram showing a changed part of the type conversion process. The following processing is performed between S1804 and S1807 in FIG. 18 instead of S1805 and S1806. As a column to be mutated, a node type / gene pair column, a node type column, and a gene column are determined (S4301). Also in this case, the column is determined irregularly based on random numbers, etc., and the column of the alternative node type and alternative gene pair, the column of the alternative node type, or the column of the alternative gene to be the replacement column is specified. (S4302), the mutation range column is rewritten to an alternative column (S4303). The alternative node type is also generated irregularly based on a random number or the like.

  37, the node type of the third node is replaced from the terminal (t) to the non-terminal (n), and the node type of the eighth node is terminated from the non-terminal (n). An example of replacing with (t) will be described. Note that the number of elements in the column to be mutated is one or more, and there is a case where an element alone, that is, one node type is replaced.

  In this way, the phenotype individual structure of the individual ho with the node type replaced is as shown in FIG. FIG. 45 shows an image of this phenotype. When this image is compared with the phenotypic image of the individual shown in FIG. 41, the upper part structure (4101) and the lower part structure (4102) that were in a serial relationship in the current generation, It can be seen that both are connected in parallel to the third node. In this example, an example of shifting from a serial relationship to a parallel relationship has been shown. However, if the node type is changed in reverse, it is naturally possible to shift from a parallel relationship to a serial relationship. As described above, when the node type changes, the arrangement relationship dynamically changes in parts structure (partial tree) units.

In the conventional reproductive process, the replacement of the gene itself as a node is the main operation, and the part structure placement operation due to the change of the node type is not performed.
Embodiment 5. FIG.
In the present embodiment, a form in which the diversity of reproduction is further enhanced based on the above-described technique will be described. For this reason, a matrix type individual structure, which is a developmental aspect of the genotype individual structure, is adopted. In the matrix type individual structure according to this example, each record has a plurality of gene candidates, and one of the genes is selectively reflected in the phenotype individual structure. The determination of termination and non-termination is also determined in consideration of instruction items other than the node type item.

  FIG. 46 is a diagram showing a configuration of the evolutionary computation system according to the fifth embodiment. Instead of the above-described genotype individual structure, it operates using a matrix type individual structure.

  The matrix type individual structure will be described. FIG. 47 is a diagram illustrating an example of a matrix type individual structure. For each record (corresponding to gene information), node type, gene for non-terminal node, parameter for non-terminal node, gene for terminal node, first connection destination termination instruction, second connection destination termination instruction, first connection destination termination A gene and a second connection destination termination gene are stored. Hereinafter, the specification of this structure is demonstrated easily. In the first record, when the node type is non-terminal (n), the non-terminal node gene and the non-terminal node parameter are adopted, and when the node type is terminal (t), the terminal node gene is adopted. The number of linked genes is obtained from the gene table in the same manner as described above, and the link destination termination instruction is determined by the number of linked numbers. When the contact end instruction is ON, the connection destination is the end node, and the connection end gene of the subsequent item is adopted. When the contact termination instruction is OFF, the terminal type of the next record is determined according to the node type. In this case, as in the case of the first record, a non-terminal node gene and a non-terminal node parameter, or a terminal node gene is employed.

  In the upper matrix of FIG. 47, values (logical values) that directly indicate the definition of each item described above are described. However, in the actual matrix type individual structure, the definition of each item is as in the lower matrix of the figure. A value (actual value) that does not necessarily follow is stored. These actual values are used after being interpreted as logical values in accordance with the definition of the item in which the values are stored. In the case of a node type, if the actual value is an odd number, it is treated as a terminal (t), and if it is an even number, it is treated as a non-terminal (n). If it is a gene item, that is, a non-terminal node gene, a terminal node gene, a first connection destination termination gene, and a second connection destination termination gene, the number of identified genes (in this example, 6 from GL) The gene is specified by the remainder. In this example, a numerical value with a remainder of 1 is treated as “G”. Similarly, in the case of 2, “H”, in the case of 3, “I”, in the case of 4, “J”, and in the case of 5 “K” is treated as “L” when 0. In the case of a parameter for a non-terminal node, the actual value is replaced with the effective range of the parameter. In this example, since three values “1”, “2”, or “3” are valid values, the parameter is specified by a remainder of 3. A numerical value with a remainder of 1 is treated as “1”. Similarly, a value of 2 is treated as “2”, and a value of 0 is treated as “3”. In the case of the items of the first connection destination termination instruction and the second connection destination termination instruction, when the actual value is an odd number, it is treated as “ON”, and when it is an even number, it is treated as “OFF”.

  In this example, the interpretation method of converting the actual value to the logical value by the remainder of the number of logical values that can be taken for each item with respect to the actual value has been described, but it may be interpreted by other methods. However, a method in which the probability of conversion to each logical value is the same is desirable. For example, there is a method in which the setting range of the actual value is limited to 1 to 12, divided into layers of the number of logical values, and a corresponding logical value is assigned to each layer. For example, the actual value number 12 is divided by the parameter effective value number 3 to obtain the size of the layer as 4, and the actual value ranges from 1 to 4 layers, 5 to 8 layers, and 9 to 12 layers. If the actual value is 1 to 4, it is treated as parameter “1”, if the actual value is 5 to 8, it is treated as parameter “2”, and if the actual value is 9 to 12, it is treated as parameter “3”.

  When setting the actual value setting range, it is effective to use the least common multiple (or an integer multiple of the least common multiple) of the number of logical values that can be taken for each item in order to keep the expected value of each logical value constant. It is.

  FIG. 48 is a diagram illustrating a configuration of the fitness calculation unit. Instead of the genotype individual structure, the matrix type individual structure is converted into a phenotype individual structure by the type conversion unit 501. The phenotype individual structure is the same as that of the above-described embodiment.

  The type conversion process by the type conversion unit 501 will be described. FIG. 49 is a diagram illustrating the configuration of the type conversion unit. The type conversion unit 501 includes a matrix type individual structure storage unit 4901, a route processing unit 4902, a record reading unit 4903, a terminal node setting unit 4904, a non-terminal node setting unit 4905, a phenotype solid structure storage unit 4906, a node No increment unit 4907, A connection position determination unit 4908 and a gene determination unit 4909 are included.

  FIG. 50 is a diagram showing a flow of type conversion processing. In route processing (S5001) by the route processing unit 4902, the node of the route is determined. Thereafter, the following processing is repeated for each node in order from the root (S5002), and the loop processing is repeated until there is no connection position in the connection position determination processing (S5003) by the connection position determination unit 4908. In the gene determination process (S5004) by the gene determination unit 4909, the node type, gene, and the like are determined (S5005). When the node type is the termination (t), the termination node setting process by the termination node setting unit 4904 (S5006) In the case of non-terminal (n), setting as a non-terminal node is performed in the non-terminal node setting process (S5007) by the non-terminal node setting unit 4905. Hereinafter, each process is explained in full detail.

  First, route processing (S5001) will be described. FIG. 51 is a diagram showing a flow of route processing. First, record reading processing from the matrix type individual structure is performed (S5101). Specifically, the record reading unit 4903 is instructed to read, and a record is acquired. The record reading unit 4903 sequentially reads records from the matrix type individual structure and delivers them each time there is a read instruction. Further, it is configured to return the No of the latest delivered record in response to the inquiry. Here, since the first read instruction is received, the first record is delivered to the route processing unit 4902. Next, the route processing unit 4902 determines the node type of the read record (S5102). In the case of the end (t), the end (t) is output as the node type, and the end of the read record as a gene. The node gene is output (S5103). On the other hand, in the case of non-terminal (n), non-terminal (n) is output as the node type, the gene for the non-terminal node of the read record is output as the gene, and the parameter for the non-terminal node of the read record is output as the parameter (S5104).

  The termination node setting process (S5006) will be described. FIG. 52 is a diagram showing a flow of terminal node setting processing. The gene output from the route processing unit 4902 or the gene determination unit 4909 is written to the record of the node of the phenotype individual structure (S5201), and 0 is similarly written to the connection number of the record of the node (S5202).

  The non-terminal node setting process (S5007) will be described. FIG. 53 is a diagram showing a flow of non-terminal node setting processing. The number of linkages corresponding to the gene output from the route processing unit 4902 or the gene determination unit 4909 is acquired from the gene table (S5301). Then, the gene, the parameter, and the number of connections are written in the record of the node of the phenotype individual structure (S5302). Further, the reservation code is written in the connection destination node numbers corresponding to the number of connections included in the record of the node (S5303). In addition, the record reading unit 4903 is inquired, the latest record No. read by the record reading process is acquired, and written in the source record No. of the record of the node (S5304).

  FIG. 54 is a diagram showing a process of generating a phenotype individual structure for an individual. Since the node type of the first record is non-terminal (n), the non-terminal gene “J”, the non-terminal parameter “3”, and the number of connections “2” are written, and reservation codes are set in the two connection destination nodes No. In addition, as the record read by the record reading process, the first record is the latest at this time, so “1” is set in the source record No.

  Next, the connection position determination process (S5003) will be described. FIG. 55 is a diagram showing a flow of a connection position determination process. A reservation code to be linked is specified from the linked nodes in each record of the phenotype individual structure (S5501). If there is no reservation code (S5502), since there is no connection position, the end status is set to “none” and the process ends (S5505). If there is a reservation code (S5502), the reservation code is rewritten to No of the node (S5503), and the end status is set to “Yes” and the process ends (S5504).

  Next, the gene determination process (S5004) will be described. 56 and 57 are diagrams showing a flow of gene determination processing. First, the source record No. included in the record of the reservation code to be linked is read (S5601). When determining the gene of the second node in FIG. 54, the record of the reservation code that is the connection target is the record of the first node, and the source record No. included in the record is “1”. Next, the source record corresponding to the source record No of the matrix type individual structure is specified (S5602). In the example of the second node, the first record of the matrix type individual structure of FIG. 47 is specified. Next, the contact termination instruction is read from the source record. At this time, if the reservation code to be connected is the first connection destination node No, the first contact end instruction is read, and if the reservation code is the second connection destination node No, the second contact end instruction is read. Is read (S5603). In the example of the second node, as shown in FIG. 54, since it is connected to the first connection destination node No in which the reservation code is set, the first contact end instruction “ON” of the first record of the matrix type individual structure Is read.

  If the read contact end instruction is “ON” (S5604), the contact end node gene is read from the source record. At this time, if it is determined by the first contact termination instruction, the first contact termination node gene is read. If it is determined by the second contact termination instruction, the second contact termination node gene is read (S5605). ). Then, the terminal type (t) is output as the node type, and the contact terminal node gene is output as the gene (S5606). In the example of the second node, the end (t) and the first connection destination end node gene “G” are output.

  On the other hand, if the read contact termination instruction is “OFF” (S5604), the next record is acquired by the record reading process (S5607) from the matrix type individual structure. If the node type of the read record is the end (t) (S5608), the end (t) is output as the node type, and the end node gene of the read record is output as the gene (S5609). When the node type of the read record is non-terminal (n) (S5608), the non-terminal (n) is output as the node type, the gene for the non-terminal node of the read record is output as the gene, and is read as a parameter The non-terminal node parameter of the record is output (S5610).

  In the matrix-type individual structure of FIG. 47, the shaded portion indicates the elements used in the gene determination process (S5004). The other parts are not used for determination.

  In this way, the phenotype individual structure for the individual shown in FIG. 58 is obtained. FIG. 59 is a diagram showing a phenotype image of an individual.

  The crossover process (S1202) and the mutation process (S1203) included in the reproductive process shown in FIG. 12 are different from the above-described embodiment.

  FIG. 60 is a diagram showing a flow of crossover processing. In the crossover process, blocks of the same size (6103, 6104) are exchanged as in the crossover process image shown in FIG. The block is a part of the matrix (6101, 6102) and is at least the size of the smaller matrix.

  The process of exchanging blocks is repeated until a predetermined number of individuals is obtained (S6001). Two individuals to be crossed are specified (S6002), and the size of the cross block is specified (S6003). The size of the crossing block is represented by the number of rows and the number of columns (the number of records and the number of items). The number of rows in the block is in the range of 1 or more and less than the number of rows in the matrix of the two specified individuals (the number of small side rows), and the number of columns in the block is 1 or more and the number of columns in the matrix In the following ranges, each is determined randomly using random numbers or the like. Next, a cross block is specified for each of the two individuals according to this size (S6004). Specifically, the difference row number is obtained by subtracting the block row number from the minor row number, and the block start row is determined irregularly using random numbers or the like in the range of 1 to the difference row number + 1. Further, the number of difference columns is obtained by subtracting the number of columns of the block from the number of columns of the matrix, and the block start sequence is determined irregularly using random numbers or the like in the range of 1 to the number of difference columns + 1. Then, the crossing block is exchanged (S6005). Specifically, the blocks composed of the number of rows from the block start row and the number of columns from the block start column are exchanged. An individual obtained by crossover from the exchanged matrix is specified (S6006). In general, both are left for the next generation, but only one may be selected and left. Finally, the process ends when a predetermined number of individuals is obtained (S6007).

  FIG. 62 is a diagram showing a flow of mutation processing. In the mutation process, the block (6303) in the matrix (6301) is replaced with an arbitrary substitute block (6302) as in the image of the mutation process shown in FIG. The block (6303) is a part of the matrix (6301) and is at least smaller than the matrix.

  The process of replacing blocks is repeated until a predetermined number of individuals is obtained (S6201). First, an individual to be mutated is identified (S6202). Next, the size of the mutation block is specified (S6203). The size of the mutation block is represented by the number of rows and the number of columns (the number of records and the number of items). The number of rows in the block is in the range of 1 or more and less than or equal to the number of rows of the specified individual matrix, and the number of columns in the block is in the range of 1 or more and less than or equal to the number of columns in the matrix. Decide on rules. Next, an individual mutation block is identified (S6204). Specifically, the number of differential rows is obtained by subtracting the number of rows of blocks from the number of rows of the matrix, and the block start rows are determined irregularly using random numbers or the like in the range of 1 to the number of differential rows + 1. Further, the number of difference columns is obtained by subtracting the number of columns of the block from the number of columns of the matrix, and the block start sequence is determined irregularly using random numbers or the like in the range of 1 to the number of difference columns + 1. Next, an alternative block to be replaced is specified (S6205). The replacement block is the same size as the mutation block, and the values in the valid range are randomly obtained by using random numbers or the like and randomly arranged. Then, the mutated block is replaced with a substitute block (S6206). Finally, the process is terminated when a predetermined number of individuals is obtained (S6207).

Embodiment 6 FIG.
In the fifth embodiment, the size of each element, that is, the length of each item is made the same, and in the crossover process and the mutation process, a block is set for each element. In the present embodiment, an example will be described in which each item length is determined separately and a block in bit units is subjected to crossover processing and mutation processing.

  For example, the node type is 1 bit, the non-terminal node gene is 4 bits, the non-terminal node parameter is 2 bits, the terminal node gene is 4 bits, the first connection destination termination instruction and the second connection destination termination instruction. 1 bit each, 4 bits are assigned to each of the first linkage destination termination gene and the second linkage destination termination gene, and the matrix type individual structure is described with a record length of 21 bits. Then, a logical value is specified by a bit structure for each item.

  The crossover process shown in FIG. 60 will be described. In specifying the crossing block size (S6003), the size is represented by the number of rows (number of records) and the number of bits. The number of bits of the block is irregularly determined using random numbers or the like within a range of 1 or more and less than the number of bits of the matrix. In specifying the crossing block (S6004). The number of bits of the block is subtracted from the number of bits of the matrix to obtain the number of difference bits, and the block start bit is determined irregularly using random numbers or the like in the range of 1 to the number of difference bits + 1. In the exchange of crossing blocks (S6005), the blocks composed of the number of rows from the block start row and the bit string from the block start bit to the number of bits of the block are exchanged.

The mutation process shown in FIG. 62 will be described. In specifying the mutation block size (S6203), the size is represented by the number of rows (number of records) and the number of bits. The number of bits of the block is irregularly determined using random numbers or the like within a range of 1 or more and less than the number of bits of the matrix. In specifying the mutation block (S6204), the number of bits of the block is subtracted from the number of bits of the matrix to obtain the difference bit number, and the block start bit is irregularly set using a random number or the like in the range of 1 to the difference bit number + 1. Decide. In specifying the replacement block (S6205),
Each bit is determined irregularly using a random number or the like.

  Although an example of a rectangular area is shown as a block, the block may not be a rectangular area. The shape is arbitrary, and it is sufficient if the blocks to be crossed have the same size and the same shape. It is characterized in that it spans multiple records.

  In the present invention, the structure and numerical values can be optimized simultaneously, and the phenotype is made compact. This is because functions that have been realized by connecting functions in the past can be implemented by a single function using parameters.

  In the case of conventional evolutionary computation, all genetic information of an individual is replaced with a phenotype regardless of whether it is a genetic algorithm or genetic programming. In the present invention, information that specifies a phenotype is retained. In addition to the portion (referred to as “exon”), the portion that does not affect the phenotype (referred to as “intron”) is included. As a result, potential qualities can be passed on to the next generation.

  The evolutionary computing system is a computer, and each element can execute processing by a program. Further, the program can be stored in a storage medium so that the computer can read the program from the storage medium.

It is a figure which shows the structure of an evolution calculation system. It is a figure which shows the example of a gene table. It is a figure which shows the flow of the learning process by an evolution calculation algorithm. It is a figure which shows the genotype individual structure of individual i. It is a figure which shows the structure of a fitness calculation part. It is a figure which shows the flow of a fitness calculation process. It is a figure which shows the flow of a type conversion process. It is a figure which shows the production | generation process of the phenotype individual structure of individual i. It is a figure which shows the phenotype individual structure of individual i. It is a figure which shows the image of the phenotype of individual i. It is a figure which shows the flow of an individual calculation process. It is a figure which shows the flow of a reproductive process. It is a figure which shows the flow of a crossover process. It is a figure which shows the genotype individual structure of an individual. It is a figure which shows the image of the phenotype of an individual. It is a figure which shows the genotype individual structure of individual C. It is a figure which shows the image of the phenotype of individual C. It is a figure which shows the flow of a mutation process. It is a figure which shows the flow of a life extension process. It is a figure which shows S type | formula of individual Lisp. It is a figure which shows S type | formula of Lisp of an individual. It is a figure which shows S type | formula of Lisp which crossed. It is a figure which shows S type | formula of Lisp of an individual C. It is a figure which shows the genotype individual structure of the individual (a) which added the parameter. It is a figure which shows the change part of a type conversion process. It is a figure which shows the phenotype individual structure of the individual (a) which added the parameter. It is a figure which shows the image of the phenotype of the individual i which added the parameter. It is a figure which shows the change part of an individual calculation process. It is a figure which shows the change part of a crossover process. It is a figure which shows the genotype individual structure of the individual | tissue which added the parameter. It is a figure which shows the genotype individual structure of the individual C which added the parameter. It is a figure which shows the image of the phenotype of the individual C which added the parameter. It is a figure which shows the changed part of a mutation process. It is a figure which shows the change part of a crossover process. It is a figure which shows the changed part of a mutation process. It is a figure which shows the example of a gene table. It is a figure which shows the genotype individual structure of the individual e. It is a figure which shows the change part of a type conversion process. It is a figure which shows the production | generation process of the phenotype individual structure of an individual. It is a figure which shows the phenotype individual structure of an individual D. It is a figure which shows the image of the phenotype of an individual D. It is a figure which shows the change part of a crossover process. It is a figure which shows the change part of a type conversion process. It is a figure which shows the phenotype individual structure of the individual e. It is a figure which shows the image of the phenotype of individual ho. FIG. 10 is a diagram illustrating a configuration of an evolutionary computation system according to a fifth embodiment. It is a figure which shows the example of a matrix type individual structure. It is a figure which shows the structure of a fitness calculation part. It is a figure which shows the structure of a type | mold conversion part. It is a figure which shows the flow of a type conversion process. It is a figure which shows the flow of a route process. It is a figure which shows the flow of a termination | terminus node setting process. It is a figure which shows the flow of a non-terminal node setting process. It is a figure which shows the production | generation process of the phenotype individual structure to an individual. It is a figure which shows the flow of a connection position determination process. It is a figure which shows the flow (1/2) of a gene determination process. It is a figure which shows the flow (2/2) of a gene determination process. It is a figure which shows the phenotype individual structure to an individual. It is a figure which shows the image of the phenotype to an individual | organism | solid. It is a figure which shows the flow of a crossover process. It is a figure which shows the image of a crossover process. It is a figure which shows the flow of a mutation process. It is a figure which shows the image of a mutation process.

Explanation of symbols

101 initial individual input unit, 102 individual group storage unit, 103 fitness calculation unit, 104 gene table, 105 fitness storage unit, 106 individual selection unit, 107 selected individual storage unit, 108 reproductive unit, 109 learning individual output unit, 501 Type conversion unit, 502 Individual operation unit, 503 Data comparison unit, 4901 Matrix type individual structure storage unit, 4902 Route processing unit, 4903 Record reading unit, 4904 Terminal node setting unit, 4905 Non-terminal node setting unit, 4906 Expression type solid structure storage Part, 4907 node No increment part, 4908 connection position determination part, 4909 gene determination part.

Claims (8)

An evolutionary computation system that performs evolutionary computation using genotypic individual structure information in which genetic information including a gene that identifies a function is continuous, and has the following elements: (1) In the same generation An individual population storage unit (2) that stores genotype individual structure information of each individual that is an element of the individual population. The number of data input by the function specified by the gene is defined as the number of connections. Gene table (3) Sequentially reads genes included in genotype individual structure information, obtains the number of linked genes from the gene table, identifies the genes to be linked nodes to satisfy the linked number, and Generates phenotypic individual structure information that describes connectivity, inputs learning input data, and is identified by phenotypic individual structure information Adaptation that calculates individual fitness output data by comparing individual computation output data with learning target data by sequentially performing computations with functions specified by each gene according to the gene connection relationship. Degree calculation unit (4) individual selection unit for selecting an individual from an individual population based on fitness (5) part of information is exchanged between a plurality of individuals for the genotype individual structure information of the selected individual A reproductive unit that performs crossover processing or mutation processing that changes some information, obtains genotype individual structure information of each individual that is an element of the next generation individual population, and stores it in the individual population storage unit. The genetic information of the genotype individual structure information further includes parameters added to the gene,
The evolution calculation system according to claim 1, wherein the fitness calculation unit executes the calculation using a parameter added to the gene when performing the calculation based on the function specified by the gene. The genetic information of the genotype individual structure information further includes a node type indicating whether the gene is a terminal node or a non-terminal node when the gene is a node constituting a connection relationship,
When generating the phenotype individual structure information describing the linkage relationship between genes, the fitness calculation unit sets the number of linked genes to 0 when the node type of the gene indicates the end node. The evolution calculation system according to claim 1, wherein: The genetic information of the genotype individual structure information indicates at least a gene, a parameter added to the gene, and whether it becomes a terminal node or a non-terminal node when the gene becomes a node constituting a connection relationship. It is a record that includes an item of node type, the genotype individual structure information is matrix type individual structure information constituting a matrix composed of records and items,
The reproductive part performs a crossover process for exchanging a block across a plurality of records as a part of a matrix or a mutation process for changing a block across a plurality of records as a part of a matrix. Evolutionary computing system. The genetic information of the genotype individual structure information includes at least a bit configuration indicating a gene, a bit configuration indicating a parameter added to the gene, and a terminal node or non-terminal when the gene becomes a node constituting a connection relationship It is a record including a bit configuration indicating a node type indicating whether it is a node, and the genotype individual structure information is matrix type individual structure information constituting a matrix composed of records and bits,
The reproductive part performs a crossover process for exchanging a block across a plurality of records as a part of a matrix or a mutation process for changing a block across a plurality of records as a part of a matrix. Evolutionary computing system. For each individual that is an element of an individual group related to the same generation, an individual group storage unit that stores genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and an associated gene and specified by the gene An evolution calculation method using an evolution calculation system that has a gene table that defines the number of data input by the function to be input as the number of connections and performs evolution calculation using genotype individual structure information, and has the following elements (1) The genes included in the genotype individual structure information are sequentially read, the number of linkages of the genes is obtained from the gene table, and the genes that are the nodes to be linked to satisfy the number of linkages are obtained. Identify and generate phenotypic individual structure information that describes the linkage between genes, input learning data, According to the linkage relationship of the genes specified in the report, the calculation by the function specified by each gene is sequentially executed to obtain the individual calculation output data, and the individual calculation output data is compared with the target data for learning, and the adaptation for each individual (2) Individual selection step for selecting an individual from an individual population based on fitness (3) Part of the genotype individual structure information of the selected individual among a plurality of individuals The crossover process that exchanges the information of each other, or the mutation process that changes part of the information, the genotype individual structure information of each individual that becomes the element of the next generation individual population is obtained and stored in the individual population storage unit Process. For each individual that is an element of an individual group related to the same generation, an individual group storage unit that stores genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and an associated gene and specified by the gene The computer has a gene table that defines the number of data input by the function to be input as the number of connections, and uses the genotype individual structure information to perform evolutionary calculations. Computer-readable recording medium on which the program is recorded (1) Genes included in the genotype individual structure information are sequentially read, the number of linked genes is obtained from the gene table, and becomes a connection destination node so as to satisfy the number of linkages Identify genes, generate phenotypic individual structure information that describes the linkage relationships between genes, Input the data, and sequentially execute the calculation by the function specified by each gene according to the linkage relationship of the genes specified by the phenotype individual structure information, obtain the individual calculation output data, and use the individual calculation output data as the learning target Fitness calculation processing for calculating fitness for each individual compared with data (2) individual selection processing for selecting individuals from population based on fitness (3) genotype individual structure information of selected individuals On the other hand, crossover processing for exchanging part of information between multiple individuals or mutation processing for changing part of information is performed, and genotype individual structure information of each individual that becomes an element of the next generation individual population is obtained. Reproductive processing to be obtained and stored in the individual population storage unit. For each individual that is an element of an individual group related to the same generation, an individual group storage unit that stores genotype individual structure information in which genetic information including a gene that identifies a function is continuous, and an associated gene and specified by the gene For a computer that is an evolutionary computation system that performs evolutionary computation using genotype individual structure information, and has a gene table that defines the number of data input by the function to be input as the number of connections. Program (1) Sequentially reads genes included in genotype individual structure information, obtains the number of linked genes from the gene table, specifies the genes to be linked nodes to satisfy the linked number, and links between the genes Generate phenotypic individual structure information that describes the relationship, input learning input data, and In order to calculate the fitness of each individual, the individual calculation output data is obtained by sequentially executing the calculation by the function specified by each gene according to the linkage relationship of the genes to be calculated, and the individual calculation output data is compared with the learning target data. (2) Individual selection procedure for selecting an individual from an individual population on the basis of fitness (3) Partial information among a plurality of individuals with respect to genotype individual structure information of the selected individual A reproductive procedure in which crossover processing to be exchanged or mutation processing that changes some information is performed, genotype individual structure information of each individual that is an element of the next generation individual population is obtained and stored in the individual population storage unit.

Images