Research On Neuroevolution Method Based On Bionic Optimization Algorithm

Neuroevolution plays a decisive role in the further development of neural networks.Traditional shallow neuroevolution methods mainly change the initial weights of the network.It can not maximize the accuracy of the neural network.The emergence of deep neuroevolution methods opened up a new way for the research on automatic design of neural network structure.But they still have much room for improvement in how to define the search space of network structure and how to design effective search strategies.To solve these problems,a new neuroevolution method system based on bionic optimization algorithm was proposed.This method system has been defined the weight space and structure space of the neural network from the perspective of optimization theory.The bionic optimization algorithm in the weight space and structure space has been proposed.And the method of neuroevolution for the shallow and deep neural networks has been established.In the aspect of shallow neuroevolution,an improved genetic algorithm(IGA)based on elite heuristic operation and resettlement strategy has been proposed.And an improved coyote optimization algorithm(ICOA)based on influencing weight adaptively.The shallow neuroevolution method based on IGA and ICOA has been applied to weight space of BP neural network.In the aspect of deep neuroevolution,the structure space of convolutional neural network has been proposed.It has solved the search space design of neural architecture search(NAS),and the deep neuroevolution method based on GA in the structure space of convolutional neural network has been proposed.This method has solved the problem of explosive combination of many hyper-parameters and network structure parameters when designing deep learning model.Neuroevolution method use efficient bionic optimization algorithm as the search strategy.And it automatically obtain the neural network with strong generalization ability and hardware friendliness.It can greatly reduce the cost of research and development,and liberate the labor force of researchers.It has great practical significance.The dissertation has 24 figures,8 tables and 95 references.