基于两次赌轮选择的神经网络遗传优化  

Evolving neural network with dual roulette wheel selection

作　　者：张晓文[1] 杨煜普[1] 许晓鸣[1] 

机构地区：[1]上海交通大学自动化系,上海200030

出　　处：《控制与决策》2005年第2期210-213,共4页Control and Decision

摘　　要：从神经网络输入输出映射关系的实质是一种广义级数展开的观点出发,提出利用两次赌轮法,按照基因片段的功能划分来对交叉部位进行再次选择,以使功能相似的基因片段不会出现在同一个染色体中,避免神经网络遗传优化中"近亲繁殖"的产生.一个两类分类问题的仿真分析表明,该算法非常有效,能使网络持续收敛到所能达到的最小误差,从而得到全局最优解.According to the viewpoint that mapping function of the neural network from input to output is in fact a generalized series expansion, a new crossover strategy that takes dual roulette wheel selection to choose the crossover segment according to its function is proposed. It can avoid segments with similar functions appearing in the same off-spring, and then the occurrence of the Competing Conventions Problem in genetic optimizing neural networks is prevented. Simulation on a two-class classification problem shows that such strategy is very efficient, and the network being evolved is well convergent until the best global optimum solution is found.

关 键 词：神经网络 遗传优化 两次赌轮法 树图编码 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]