Structural optimization of neural network by genetic algorithm with damaged genes
Tetsuyuki Takahama, Setsuko Sakai · 2004
There are some difficulties in researches on supervised learning using neural networks: difficulty of selection of a proper network structure, and difficulty of interpretation of the hidden units. In this paper, DGGA (Genetic Algorithm with Damaged Genes) is proposed to optimize the network structure of neural networks. DGGA employs real-coded genetic algorithm and introduces the idea of genetic damage. In DGGA, the information of damaged rate is added to each gene. DGGA inactivates the genes that have lower effectiveness using genetic damage. The performance of DGGA for structural optimization is shown by optimizing a simple problem. Also, it is shown that DGGA is an efficient algorithm for structural optimization of neural network by applying DGGA to learning of a logical function.