进化神经网络间接编码方法的计算框架研究

针对传统的直接编码方法对大规模神经网络难以进化的问题,研究者提出了进化神经网络的间接编码方法,这类方法的核心思想是网络子结构的重复可通过一组基因的多次表达来实现从基因型到表现型的映射,这种基因重用机制可在较小的基因空间中进行大规模神经网络的快速搜索.本文在总结和分析各类间接编码实现方法的基础上,给出了进化神经网络间接编码方法的一般性计算框架,每一次神经网络的进化过程分为三个阶段:发育阶段、学习阶段和进化阶段.并从计算框架的基因组进化过程和神经网络发育过程两个方面比较分析了各种典型间接编码方法的优缺点.

A Computational Framework for Indirect Encoding Methods of Evolutionary Neural Networks

According to the difficulties in the evolving large scale neural networks using the traditional direct encoding methods,many researchers are proposing the novel indirect encoding methods for evolutionary neural networks.That is,a network structure that repeats many times can be represented by a single set of genes that is reused in mapping from genotype to phenotype,and such genetic reuse allows searching the large scale neural networks through a lower dimensional genotypic space.In this paper,we introduce a general computational framework for the indirect encoding methods of evolutionary neural networks through the study of existed methods,in which every evolutionary process of neural networks is divided into three stages:development,learning and evolution.Additionally,we analyze the advantages and disadvantages for the different indirect encoding methods from two aspects of the computational framework:genome evolution and neural network development.