The Improvement Of Differential Evolution Algorithm And Its Research In Neural Network Optimization

The optimization problem has always been involved in all aspects of the engineering field and scientific inquiry.Up to now,science and technology are constantly updated,and the traditional optimization method has shown unsatisfactory results in the face of complex and huge data problems such as high dimensions and multiple constraints.The evolutionary algorithm has the advantages of simple principle,easy to operate and high robustness et.al,which provides an effective path to solve the optimal problem.As a global optimal iterator in evolutionary algorithms,the faster convergence,robustness and accuracy of differential evolution algorithm have attracted the attention of many scholars in the field of optimization.Nonetheless,differential evolution algorithms are also influenced by internal factors such as mutation variants,mutation factors,crossover factors,selection operators,as well as external factors such as the dimensions of problem space and range of decision space.Therefore,two methods are proposed to promote the coordination between internal and external factors and extend it to the optimization of neural networks.Regarding these two aspects,the main researches include:(1)A differential evolution algorithm for enhanced multiple mutation strategy and elite selection: First,the “DE / ppbest-to-ppbad / 1” operator is proposed by using the interaction between the current best and worst individuals and the average individual of the population.Secondly,combined with the traditional “DE/rand/1” operator,a multimutation strategy was established to enhance the deficiency of single mutation mode and further improve the population diversity.Moreover,the elite selection method combined with non-dominant crowding distance sorting allows the more comprehensive optimal individuals to enter into the next evolution.Finally,through the overall framework design,effectively improve the algorithm accuracy and convergence.(2)Differential evolution algorithm of potential population regeneration framework for two-stage parameter changes: Through two-stage parameters combining the population regeneration framework,potential population individuals can regenerate and re-enter the next evolution.Meanwhile,the mutation threshold AP was set to avoid the population falling into the local optima.The experimental data show that the diversity and accuracy of the algorithm are improved.(3)Expand the improved differential evolution algorithm to optimize the neural network research: We use the underlying population regeneration framework with twostage parameters and establish the hidden layer change mechanism.The experimental results show that the algorithm is used to optimize the network structure instead of training the neural network individually,which improves the efficiency and accuracy of the neural network optimization.