Research And Application Of Large Scale Multi-layer Perceptron Neural Network

Multi-layer perceptron neutral network has the ability to process and self-study. In theory, researcher has proved that when neutral network has more than two layers of hidden neutral nodes, it can stimulate non-linear functions. It makes neutral network outstanding in solving classification problem. However in practice, neutral network has inevitable shortcomings as well, such as local minimum problem during network training process, slowness of convergence speed, complexion of network structure and poor generalization performance, which needs to be overcome in practice. It is an important topic how to optimize network structure and weights during neutral network training process.Based evolution theory, evolutionary computation is looking for random optimal solution by simulating biological evolution. Intelligence and parallelism are two key characteristics of evolutionary computation. Different from traditional BP calculation, it does not require continuously differentiable. With more robustness, evolutionary computation is suitable to solve general optimization problem.Based on evolutionary algorithm, neutral network could evolve in three ways, network connection weights, network structure and learning rules. Evolutionary network provides new solutions in improving traditional with respect to network design, network performance.The author brings out a network training method of data classification, learning neat(LNEAT). LNEAT algorithm simplifies network evolution by splitting a problem into multiple sub tasks. It learns sub tasks by applying BP rule into LNEAT. The new algorithm has the advantages of BP rules and Neat calculation in topology and weight search. In addition, it overcomes the problem caused by NEAT algorithm. The experimental results show that LNEAT could generate network to solve classification and LNEAT evolutionary neutral network has better generalization ability and more accuracy. In general, LNEAT algorithm has achieved satisfactory result in voice recognition and made big progress in network training speech.