Structure Design of Neural Networks Using Genetic Algorithms
Satoshi Mizuta
Takashi Sato
Demelo Lao
Masami Ikeda
Toshio Shimizu
Department of Electronic and Information System Engineering,
Faculty of Science and Technology, Hirosaki University,
3, Bunkyo-cho, Hirosaki 036-8561, Japan
Abstract
A method for designing and training neural networks using genetic algorithms is proposed, with the aim of getting the optimal structure of the network and the optimized parameter set simultaneously. For this purpose, a fitness function depending on both the output errors and simpleness in the structure of the network is introduced. The validity of this method is checked by experiments on four logical operation problems: XOR, 6XOR, 4XOR-2AND, and 2XOR-2AND-2OR; and on two other problems: 4-bit pattern copying and an 8 x 8-encoder/decoder. It is concluded that, although this method is less powerful for disconnected networks, it is useful for connected ones.
