一种改进的进化回归神经网络系统

对基本进化回归神经网络系统作了改进．首先提出一种可切换的适应度评估函数，使得适应度函数能够始终保持对训练误差的敏感性，保证选择机制正确而有效地复制优良个体；然后针对均匀变异对个体变异力度不够的问题，引入一种变邻接长度的集中变异方式，提高系统维持种群多样性和发现优良个体的能力。结合个体适应度同种群平均适应度的关系，给出了变异步长自适应调整策略；最后利用个体之间的汉明距离，对最优个体保留策略进行了改进，限制最优个体在种群中的重复复制。仿真结果表明综合上述改进后的进化回归神经网络系统有更好的性能。

A Modified Evolving Recurrent Neural Network System

Some improvements were presented for a basic evolving recurrent neural network system. Firstly,a switchable fitness evaluating function was proposed to maintain the sensitivity of fitness function in training error,and ensure the exact and effective duplication of fine individuals by selecting operation. Next,aiming at the slight mutation intensity of uniform mutation on individuals,a concentrative mutation method with variable neighborhood length was introduced to improve the system abilities to keep population diversity and to find fine individuals. Then,based on the relationship between the fitness of individual and the mean fitness of population,an adaptive adjustment strategy of mutation step size was proposed. Finally,by applying the Hamming distance among individuals,the elitist keeping strategy was also improved to stop the fittest individual from reduplicating itself in population. Simulation results show that the evolving recurrent neural network system with all these improvements is of more excellent performance.