This paper describes a phenotype genetic operation for determining the structure and connection weights for neural networks. This technique defines the network as a two-dimensional chromosome, performs crossover using two-dimensional substructures, and application of genetic operation to both determining the network structure and training the connection weights. Using problems which partially involve Exclusive-OR logic, or training data that is partly dynamically changing, the proposed method is shown to be more effective than the standard genetic algorithms with one-point, two-point, or uniform crossover in the synthesis of neural networks that have two-dimensional building blocks.
抄録全体を表示