Feedforword Neural Network Evolutionary Learning Of Multi-Layer Topologies

Learning and evolution are two fundamental forms of adaptation. Genetic algorithm is a kind of random searching method using lives' natural selection and genetic mechanism. There has been a great interest in combining learning and evolution with artificial neural networks (ANN's) in recent years. In this paper we present a new approach for automatic topology optimization of Multi-layer Feedforword Neural Networks.They are summarized as following:Firstly, the dissertation analyses characters of severaltraditional genetic algorithms for optimization. Following this, a new method, combined Pseudo-parallelism evolution technique based on sub-population competition with parent mutation mechanism, is proposed with improved BP algorithms. In contrast to other approaches it allows that two networks with different number of units can be crossed to a new valid "child" network. Compared with genetic algorithm with sharing, it has some improvements in both converging velocity and precision. Secondly, analyzing the inadequacies of the evaluation indices for premature convergence, a novel improved adaptive mutation algorithm (AGA) is described. The calculation result of an example shows that PPGA is able to get the real-time information of population diversity during the process of evolution. Finally, We applied this algorithm to N-Parity problemsand economic forecasting , the results confirm, that optimization make sense, because the generated networks verifies its validity. The result of experiment shows that the global convergence and searching velocity are both improved.