The structure of simple neural-network has been optimized by the use of Genetic Algorithm. The neural-network is a perceptron, which has 3 outputs : logical AND, OR and XOR of 2 inputs. The evaluation function for optimization is a linear combination of the correctness, the network sizes,and an auxiliary term inducing the optimum solution. The chromosome is a vector of the link weights of the network. The genetic operators used are crossing-over and point-mutation on the parents chromosomes. Two genetic rules, haploidy and diploidy were tested. In the haploidy rule, each individual has single chromosome, and the offspring is generatd by crossing-over the parents chromosomes at a randomly chosen locus and taking one of those crossed-over chromosomes. In the diploidy rule, each individual has a couple of chromosomes, as the natural animals have, and offspring's chromosomes are generated by combining the gamete which was again produced through the meiosis of the parents chromosomes. The other model used in the genetic algorithm is the geographical isolation model, where the entire population is divided into 4 sub-populations, in which the local selection and reproduction are carried-out, though, in some time interval, randomly sampled individuals are exchanged among sub-populations. Comparison was made among 4 combinations of haploid or diploid, and single-population or multiple sub-populations. Diploidy together with sub-population model was proved to be the best for this optimization problem, here the well-known optimum structure of network was found.
抄録全体を表示