Research On Architecture Of Genetic Evolutionary Neural Networks

EANN is a combination of neural networks and genetic algorithms. Research on EANN is more active and attracted by people. There have already been many valuable results and achievement, and there have already been some successful application instance in engineering. Thus, EANN has a promising future.In the first place, the thesis presents the history,theory basis and application fields of genetic algorithms and neural networks. Secondly, the thesis gives a emphasis introduction on genetic neural networks and particularly discusses genetic neural networks' (GNN's) key techniques and encoding methods here. At the same time, the thesis presents genetic BP neural networks and employs it to solve mouse gesture recognization. Thirdly, the thesis discusses some shortcomings of current EANN methods. The thesis focuses on sructrural-functional mapping problem, damanage innovation with mutation problem and keeping neural networks size minimal problem and discusses some solutions, then proposes a new method to evolve network architecture. With this method, the three problems discussed above are avoided.The thesis proposes node-based encoding method which describes networks' structure and weights. When new nodes and links are created, the historical data is used to avoid structural-functional mapping problem. The thesis applies niching strategy to keep population diversity. Population is confirmed by using explicit fitness sharing and calculating individual phentype's Osh distance. Earliness of the population is prevented by keeping population diversity. In order to keep the size of the networks as smaller as possible, the thesis employs the simplest topology structure without hidden layer at the beginning of the evolution, and the topology is expanded by structural mutation.With these results, the experiment shows that the application of this method in obstacles avoidance path planning for robot is viable and more effective.