Research On Neural Network Structure Optimization Based On Evolution Algorithm

Construction of neural network impacts neural network' s learning ability and geralization ability. The simpler the network structure is, the more insufficient the learning ability becomes. Vice versa. The structure optimaztion is just to make the neural network get a proper structure which satisfied both its learning ability and generalization ability. Evolutionary computation simulated biologic evolution process is a good optimation method.The object of this research is function approximation BP neural network. There are three aspects in this paper:Researching for Neural networks learning method; researching for function complexity;and researching for neural networks structure optimation based on evolutionary computation.The first of this research is about some methods of neural networks's constructon optimation. Three limitions of BP networks have been presented. To improve these limitions, LMBP algorithm is presented. This paper improved LMBP algorithm, by rules of error changing based on quadratic error and gradient reducing, analyzing the optimization of the network's structure and adding hidden neurons or adding network layers one by one adaptively, as a result a proper structure of the network is got.The second section included this research is about function complexity. Analyzing the great deal of data of improved LMBP algorithm, the following conclusion has been drawn: the more inflexions in a function, the bigger structure that approximating the function would be need. A simple estimation of ruction complexity in this paper is presented. Based on experience data, this method can estimate neural network initial structure according to the count of the function's inflexions.The research on nural networks construction optimal based on evolutionalgorithm is included in the third section. Among evolution algorithms,the characters of Co-evolving parasites genetic algorithm are suitednetworks construction optimal. Organised some individuals of a samestructure as a sub-population, and organized some sub-population as a bigpolulation. Operating of Selection, changing, crossover and mutation inthe inner of sub-populations, operating of changing information in thebig population. So the weights and the structure can be evoluted in thesame time. In this way, a new co-evolving parasites genetic algorithm hasbeen designed. Mixed the improved LMBP algorithm, Simulation experienceand the analyzing of this simulation result are provided and the furtherwork direction are proposed.