进化编程优化RBF神经网络的结构和参数

本文利用进化编程（ＦＰ）来同时进化径向基函数神经网络（ＲＢＦＮＮ）的结构和参数。与其它进化神经网络方法有以下四个方面的不同：（１）ＥＰ是基于拉马克的进化学说，强调父代与子代之间的行为联接；（２）进化算子中仅有突变，而没有交叉，以消除互换问题；（３）突变操作中，删除总是先于添加进行，以获得最简的网络结构；（４）利用测试样本集构造适应度函数，以提高网络的泛化能力。用进化ＲＢＦＮＮ来预测Ｍａｃｋｅｙ－Ｇ

OPTIMIZE THE STRUCTURES AND PARAMETERS OF RBF NEURAL NETWORKS USING EVOLUTIONARY PROGRAMMING

In this paper, the structures and parameters of radial basis function neural networks are evolved simultaneously by evolutionary programming(EP). There are four difference aspects between RBFNNs evolved by EP with other evolved neural networks: (1)EP based on Lamarckian evolution is put emphasis on the behavioral link between parents and their offspring; (2)In order to reduce the detrimental effect of the permutation problem, EP algorithm, which use mutation, rather than crossover, is adopted; (3)Node deletions are always attempted before node additions in the mutations in order to encourage the evolution of small RBFNNs; (4)The fitness function is determined through a test set. Such use of a test set improves the generalization ability of RBFNNs. The experimental results on evolved RBFNN predicting the Mackey Glass time series shows the new algorithm is very effective.