Research On Multi-population Cooperative Coevolutionary Dynamic Multiobjective Algorithm Based On Decision Variable Classification

Dynamic multiobjective optimization problems(DMOPs)not only have conflicting objects but also have the characteristic that the Pareto optimal front and/or Pareto optimal set are time-variant.Algorithm research often depends on the problem being studied.As for dynamic problems,it requires the algorithm to quickly search for the optimal solution or Pareto Front after the change,while maintaining the convergence and distribution of the solution and keep both in a balance.To solve DMOPs,this paper proposes a dynamic multiobjective algorithm based on decision variable classification.This algorithm uses the information of decision variable classification to guide search process and proposes a new cooperative coevolutionary method.Firstly,the problem is analyzed by random sampling,after that the decision space is divided into convergence control dimensions and diversity control dimensions.Then,multiple populations are used to search the PF,and these populations are divided into two categories: convergence related subpopulation and diversity related subpopulation.The number of convergence related subpopulations corresponds to the number of convergence dimensions,and each convergence related subpopulation optimizes one unique convergence dimension.A subpopulation is used to correspond to all diversity dimensions,but this subpopulation optimizes all decision dimensions.Convergence related subpopulations are carried out in a coevolutionary manner,and the values of the diversity dimensions of individuals within each subpopulation are the same.Therefore,the influence of diversity on convergence is avoided,so that convergence related subpopulations can concentrate resources to optimize individual convergence and accelerate the convergence speed of the algorithm.In order to avoid the cooperative coevolutionary algorithm's poor performance in diversity,this algorithm will selectively merge the convergence related subpopulations into diversity related subpopulation according to the degree of optimization of the convergence related subpopulations.As the algorithm proceed,all the convergence related subpopulations will gradually merge into the diversity related subpopulation,so that the diversity of the population can be improved on the premise of guaranteed population convergence.In addition,this paper also proposes an improved center point prediction strategy to respond to environment change.The strategy introduces decision variable classification information to guide the prediction process.The algorithm is tested on 16 benchmark DMOPs with different dynamic characteristics and difficulties.Compared with five representative algorithms,experiments show that the proposed algorithm can maintain the convergence and distribution of the population and respond to environment changes quickly.