基于免疫进化策略的神经网络优化方法

Optimization Algorithm Based on Immune Evolutionary Strategy for Neural Network

对神经网络的研究多年来主要集中于网络权值优化或结构优化上,却忽略了神经网络结构与权值之间密不可分的联系.针对上述问题,将免疫系统中的浓度机制和记忆机制引入进化策略,提出了一种基于免疫进化策略的神经进化算法,在优化网络拓扑结构的同时优化网络的连接权值.进一步地,用Cauchy变异算子代替传统的Gauss变异算子,以获得更为理想的全局收敛效果.理论分析和仿真结果表明,免疫进化策略能够很好地保持种群多样性,避免未成熟收敛,采用免疫进化策略设计神经网络具有良好的全局收敛性能和快速学习网络结构和网络权值的能力.

The researches on neural network ignored the closely related connection between its architecture and the weighted value for years, but mainly focused on their optimization. A neural evolution algorithm based on immune-evolution strategy is therefore proposed introducing the concentration/memory mechanism of immune system into evolutionary strategy, which can optimize simultaneously both the topological structure of network and weighted value for connection. Furthermore, the algorithm introduces the Cauchy operator instead of Gauss operator to obtain better global convergence. Theoretical analysis and simulation results both show that the algorithm can retain the population diversity, and avoid premature convergence with better global convergence as well as the ability to learn faster to learn the architecture and the weighted value of network.