Research On Multi-objective Evolutionary Algorithms Driven By Adaptive Learning Mechanism

Under the background of big data,the urgent need for efficient processing of massive,high-dimensional and uncertain data challenges the traditional information processing technology.Intelligent computing,represented by evolutionary computing,is considered to be an effective way to deal with complex multi-objective problems,and has become a research frontier and hot area in recent years.At present,multi-objective evolutionary algorithm has been widely used in many fields,and has solved many valuable practical problems,and its research results have penetrated into many disciplines.However,in the process of evolutionary multi-objective solution,we still need to focus on the following two aspects:1)How to improve the generality of the multi-objective evolutionary algorithm;2)how to integrate the characteristics of the problem into the search process of the multi-objective evolutionary algorithm,so as to achieve the efficient solution of the problem.In view of this,this paper aims to study the multi-objective evolutionary algorithm driven by adaptive learning model and its efficient solution to complex multi-objective optimization problems.The main innovative work is as follows:1)Multi-objective differential evolution algorithm and regular-based multi-objective distribution estimation algorithm have their own advantages and disadvantages for solving different types of problems.How to effectively integrate their advantages is an important way to improve the efficiency of the algorithm.Therefore,an adaptive covariance learning model driven multi-objective hybrid difference distribution estimation algorithm is proposed.Firstly,the mathematical characteristics of differential evolution operator and its influence on different types of optimization problems are analyzed by matrix theory.Secondly,the covariance matrix is used to identify the data association characteristics of population distribution,so as to construct the feature coordinate system,and the sigmoid function is used to realize the collaborative search of binomial crossover operators in different coordinate systems.Then,in the later stage of evolution,the rule model and negative correlation selection are used to make the algorithm cover the whole Pareto structure as much as possible,so as to improve the calculation eff-iciency.Finally,the algorithm is compared with three multi-objective differential evolution algorithms and three regular-based multi-objective distribution estimation algorithms on two different types of test functions.Experimental results show that the proposed algorithm can effectively solve different types of multi-objective problems and has stronger robustness.2)In order to solve the problem of low computational efficiency of inverse learning model in the multi-objective optimization problem of irregular Pareto front,an adaptive inverse learning model driven multi-objective evolutionary algorithm is proposed.The algorithm divides the whole evolution process into two stages:exploration and exploitation.In the exploration stage,the uniformly distributed reference vector is used to improve the exploration ability of the algorithm.In the exploitation stage,the non-dominated solution in the external elite archive is used to adjust the distribution of the reference vector adaptively,which is conducive to improving the exploration ability of the algorithm.In addition,the collaborative search of preference crossover and inverse learning model further improves the computational efficiency of the algorithm.Finally,it compares with six regular-based multi-objective evolutionary algorithms on 18 irregular test functions.Experimental results show that the proposed algorithm can effectively solve various types of irregular multi-objective optimization problems.3)In order to solve the problem of unbalanced distribution of evolutionary population in decision space and target space based on rule-based learning model and inverse learning model,an adaptive dual space learning model driven multi-objective evolutionary algorithm is proposed.Firstly,the algorithm introduces a method of population initialization based on serialization to identify the fitness characteristics of distance function,which is helpful to reduce the risk of the algorithm falling into Pareto’s local optimization;secondly,an adaptive mechanism is designed to adjust the allocation of computing resources,so as to realize the complementary advantages of the two learning models;secondly,the algorithm realizes the simultaneous and simultaneous optimization by integrating two space environment selection strategies At last,it compares with six regular-based multi-objective evolutionary algorithms in 22 test functions.Experimental results show that the performance of the proposed algorithm is significantly better than other comparison algorithms.4)Aiming at the modeling problems in chaotic time series forecasting and power load forecasting,a multi-objective guided sparse depth belief network is constructed.In the network training,the model transforms the single objective function with sparse penalty factor into the double objective function with reconstruction error and sparse degree,and combines multi-objective evolutionary algorithm and CD-1 method to realize automatic parameter selection,avoiding the problem of manually adjusting sparse penalty factor.The results of numerical experiments show that the multi-objective guided deep belief network is superior to other common models in the application of time series prediction.