连续参数小波神经网络的递阶遗传训练方法

Hierarchical genetic algorithm-based training of wavelet neural networks with continuous parameters

提出递阶遗传训练方法用于训练连续参数小波神经网络的参数及其结构。现有的连续参数小波网络训练方法大多只能训练网络的参数,包括平移参数、伸缩参数和权值,而网络的结构得预先用某种方法确定。应用递阶遗传算法能够把网络的结构和参数同时通过训练确定。利用混沌时间序列数据进行仿真,结果证明该模型具有较高的预测精度,提出的方法是可行的。

A hierarchical genetic algorithm is proposed to train the parameters and structure of wavelet neu- ral networks with continuous parameters. Existing training methods for wavelet neural networks with continu- ous parameters are usually confined to parameters, including connection weights, expansion parameters and movement parameters, while their structures have to be predetermined through some method. In contrast, the configuration and related parameters of wavelet neural networks with continuous parameters can be determined simultaneously by using the hierarchical genetic algorithm. A case-study, based on the chaotic time series data, illustrates the effectiveness of the proposed algorithm.