Research And Application Of Evolutionary Neural Networks

With the development of artificial intelligence technology,neural networks and evolutionary algorithms,as two important branches of artificial intelligence,have been deeply applied in various fields.The neural network has the ability of parallel computing and can deal with highly nonlinear problems,so the neural network has strong learning ability.However,it still has problems.The model structure is difficult to select,the hyperparameters are difficult to determine,and it is easy to fall into local optimum.It takes a lot of manpower and material resources to build.The evolutionary algorithm is an efficient,parallel,global search method,which can automatically acquire and accumulate knowledge about the search space in the search process,and adaptively control the search process to obtain the best solution.Therefore,scholars combined the advantages of evolutionary algorithms and neural networks to propose evolutionary neural network algorithms,which can effectively solve many shortcomings of neural networks.In this thesis,the details of evolutionary neural network algorithm are analyzed and optimized.And apply it to many fields such as image recognition,text sentiment analysis,game AI,etc.The main work has the following two aspects:On the one hand,this thesis analyzes and improves the two weight and topology evolving methods of evolutionary neural networks.Firstly,the weight mutation algorithm of evolutionary neural network based on augmented topologies is improved to make it more controllable,which can improve the search efficiency of the whole population.Secondly,this thesis combs the evolutionary neural network algorithm based on deep learning.And the algorithm details such as new species protection,population initialization,fitness function construction and parent chromosome selection are optimized to make the whole algorithm process more complete and efficient.On the other hand,this thesis experiments and analyzes the application of evolutionary neural networks.Firstly,this thesis applies the evolutionary neural network algorithm based on augmented topologies to the game AI field of reinforcement learning.And the algorithm is designed and analyzed in the three aspects which are the feasibility of the algorithm,the effect of the improved algorithm,the comparison between evolutionary neural network algorithm and classical reinforcement learning algorithm.Secondly,the evolutionary neural network algorithm based on deep learning is applied to the fields of image recognition and text sentiment analysis in deep learning and select several classic data sets for experiment and analysis on the improved algorithm effect.The model constructed by the evolutionary neural network algorithm is compared with the model built by the experts in the correct rate and total amount of parameters.In summary,this thesis improves the algorithm details of the evolutionary neural network,and proves the effectiveness of the improved evolutionary neural network algorithm.It also explores the application value of the evolutionary neural network algorithm in many fields.