Research On Ensemble Arbitrary Connected Neural Network Based On Cooperative Co-evolution Optimization

Neural networks has successfully solved many practical problems in many fields.For different problems,the topology,connection weights and internal parameters required by the neural network will also be different.It is very important for the performance of neural network to select the appropriate topology for specific problems.Traditional neural network structure optimization methods are locally optimized algorithms,and the network obtained by these algorithms tend to have a locally optimal structure.Evolutionary neural network that combines evolutionary computation and neural network has been produced,which can effectively overcome the above shortcomings.Ensemble learning is also a way to improve the generalization ability of neural networks.However,most of the current ensemble learning based on evolutionary computation focuses on the selection of the base learner and strategy optimization.They usually use the trained base learner and ignore the improvement of the accuracy of the individual learner.In summary,current algorithms still need to be improved to balance the accuracy and diversity of neural networks.Aiming at the characteristics of neural network optimization and ensemble problems,a direct coding strategy suitable for co-evolution framework is used to present arbitrarily connected neural networks in this thesis,and an optimization and ensemble framework based on co-evolution are established.This thesis introduces a variety of improvement strategies in the framework of co-evolutionary algorithms to solve these problems.To balance the accuracy and diversity of the neural network,a niche-based co-evolutionary algorithm is designed to optimize the ensemble neural network for classification in this thesis.This strategy is used to improve the local search ability in the training process.It lets different niches focus on optimizing neural networks with different structures,and the algorithm is allowed to obtain a set of highly diverse candidate networks.The crossover strategy of fully connected weight matrix is adopted to avoid the destruction of the network structure in the crossover process.In addition,a two-stage fitness is also proposed to optimize the ensemble network to obtain a compact set after training.Therefore,the algorithm can obtain networks with good ability and diversity,and can obtain their combination strategy at the same time.The main research contents of this thesis are as follows:First of all,ensemble neural network optimization problems are introduced,as well as their optimization methods.Both the traditional optimization method and the optimization method based on evolutionary computing are introduced respectively,paving the way for the subsequent improvement of algorithms and strategies.Then,the ensemble neural network optimization problem is regarded as the multi-modal optimization problem.The multi-modal evolution algorithm,niche strategy and co-evolution algorithm are introduced in detail for this problem,and the realization process of the algorithm is given in the end.Secondly,to optimize the structure and weights of the neural network at the same time,this thesis designs an optimization framework based on the coevolution algorithm which can also enhance the local search ability of the algorithm.A niche strategy is integrated in the framework of co-evolution.In this way,the scope of individual cooperation objects is restricted,the stability of the network structure is improved,and the local search ability of the algorithm is further improved.Finally,through the experiment and analysis of the evolutionary neural network,this thesis verifies the effectiveness of the proposed algorithm.Through the fitness analysis of each niche,the effectiveness of the strategy is verified.Through comparison with other related neural network optimization algorithms,the effectiveness of the proposed algorithm is demonstrated on the data set.Finally,an ensemble learning based on evolutionary computation is established on the work of the third chapter.The optimization model includes-coding strategies and analysis of multi-modal network output diversity.In the end,experiments are conducted on the data set for the proposed algorithm to verify its effectiveness.Through the comparison of different strategies,analysis of changes in ensemble complexity,and comparison with other related algorithms,the effectiveness of the algorithm is further demonstrated.