关键词: 多层感知机, 遗传算法, 随机子集, 泛化性

Abstract: MLP model training has the problems of local optimum and its unsatisfactory generalization performance. Although the genetic algorithm can jump out of local optimum, it increases computational complexity. The purpose of this paper is to overcome the above shortcomings. This SSGAMLP（Small Set Genetic Algorithm Multilayer Perceptron） model combines the genetic algorithm and MLP model. The down connection weights of MLP model nodes are regarded as low to high level mapping, so each node（including the down connection weights and thresholds） can be regarded as the expression of a feature, namely gene expression of the genetic algorithm. Individual training using a random subset along with crossover and mutation of genetic algorithm, they are equivalent to introduced random factors, thus having the possibility of obtaining the unknown feature expression. This paper’s experiment is based on MNIST dataset. Its result confirms SSGAMLP model’s advantages. The model reduces the individual computational complexity, improves the generalization. To a certain extent, this model overcomes the over fitting.

Key words: Multilayer Perceptron（MLP）, genetic algorithm, random subset, generalization performance