Improvement Of Differential Evolution Algorithm And Its Application In Neural Network

Optimization problems are ubiquitous in all fields of scientific research and engineering practice.In this day and age,science and technology are developing rapidly.Traditional optimization methods cannot achieve satisfactory solutions in an acceptable duration when solving optimization problems with large-scale,high-dimensional,NP-hard,dynamic and other complex features.Due to its simple principle,global search,good robustness and adaptability,the evolutionary algorithm has received extensive attention and recognition from researchers at home and abroad in dealing with complex optimization problems.As an emerging branch of evolutionary algorithms,the differential evolution algorithm is a type of population-based stochastic parallel algorithms.In addition to the advantages shared by the above-mentioned evolutionary algorithms,the differential evolution algorithm has many unique performances such as strong search ability and few control parameters,and has made a lot of breakthroughs in the past two decades.Despite this,the differential evolution algorithm still has the following problems to be resolved urgently:(1)The choice of the mutation strategy and the setting of the control parameters have a great influence on the performance of the differential evolution algorithm.The standard differential evolution algorithm reduces the difference in information betweenindividuals in the late evolutionary stage,and limits the global search ability of the algorithm.This not only reduces the convergence speed of the algorithm,but also causes the algorithm to fall into local optimum,which in turn leads to premature convergence and search stagnation.(2)Usually,early differential evolution algorithms evaluated their performance on a standard function test set.However,with the rapid development of human demands and production practices,more complex features are presented in the actual optimization problems.The single coding form of the differential evolution algorithm limits its optimization ability and extensive application in some practical problems.Improving the algorithm according to the specific model information in the specific optimization problem can bring more satisfactory results when applying the differential evolution algorithm to solve the specific actual optimization problem.In view of the above problems,this paper studies the improvement and application of the differential evolution algorithm.The main contributions include:(1)A multi-mutation strategy of differential evolution algorithm based on fuzzy modeling is proposed.The evolution process is divided into three stages by fuzzy modeling.Different mutation strategies and parameter settings are designed for different evolution stages,and the accuracy and efficiency of the algorithm are improved.(2)An adaptive population size of differential evolution algorithm based on niching technology is proposed.First,a heuristic clustering technique is used to divide the entire population into a series of disjoint subpopulations.Second,to increase diversity,the best individuals in the subpopulation perform uniformly random motion or Brownian motion.Finally,a population size and adaptive adjustment strategy is designed.When the number of optimal solutions no longer increases,new individuals are added to the population,and at the same time,remove the poor subpopulation in subpopulations that search on the same peak.Experimental results on the standard test set illustrate that the algorithm increases population diversity and reduces computational overhead.(3)The improved differential evolution algorithm is used to optimize neural network ensemble,which generalizes the application area of the algorithm:Traditionally,each individual network in a neural network ensemble is individually trained,which leads to network redundancy and expensive training overhead.In this paper,based on niching technology,an improved differential evolution algorithm isused to train neural network ensemble.The experimental results show that the algorithm can train all individual neural networks in neural network ensemble at the same time,and these individual neural networks are different from each other,which ensures the diversity and accuracy of individual neural networks.