基于进化规划的多层前向网络结构优化

STRUCTURAL OPTIMIZATION OF MULTILAYER FEEDFORWARD NETWORKS BASED ON EVOLUTIONARY PROGRAMMING

基于进化规划(EP)方法,该文提出了设计多层前向网络拓扑结构和权值分布的一种新算法—EPANN算法。EPANN算法能同时进化网络的结构和连接权值(包括阈值),在进化过程中,强调父代和子代之间的行为联结,结构变异既有结点删除,又有结点增加,不同于单纯的删除算法或构造算法,且结点删除总是先于结点增加,保证了网络规模尽可能小而泛化能力尽可能强。

Based on evolutionary programming, a novel algorithm named EPANN for designing the topology and weight distributions of feedforward networks is proposed in this paper. EPANN algorithm evolves network architectures and connection weights (including biases) simultaneously and emphasizes the behavioral links between parents and their offspring in evolution, such as weights training after each architectural mutation and node splitting. Unlike the pure constructive or pruning algorithm, EPANN's architectural mutations include both node deletion and node addition, and prefer node deletion to addition in order to encourage the network architecture as compact as possible and generalization ability as good as possible.