Dynamic Multi-Objective Optimization Algorithm Based On Transfer Learning

A dynamic multi-objective optimization problem is an optimization problem containing multiple conflicting objective functions,and the objective functions,constraints or decision variables change over time.Compared with static mult-iobjective optimization problems,it is more challenging to solve dynamic multi-objective optimization problems because it requires that the method is able to track the change of the Pareto front quickly and accurately.Transfer learning is one of the methods to solve dynamic multi-objective optimization problems,but the negative transfer will greatly reduce the efficiency to solve dynamic multi-objective optimization problems.Therefore,it is very important to reduce the occurrence of negative transfer when transfer learning is used for solving dynamic multi-objective optimization problems.This thesis takes account of the influence of multiple source domains on the target domain and proposes two optimization algorithms,multi-source transfer learning and source domain selection-based transfer learning,for solving dynamic multi-objective problems.The main contributions are given as follows:(1)A dynamic multi-objective optimization algorithm based on multisource transfer learning(MTL-DMOEA).In the proposed method,the feature of correlation between the source and target domains is extracted and assigned a weight to each source domain.After that,the training of the target domain will be implemented and assisted by multiple source domains.The experimental results show that the algorithm can prevent the occurrence of negative migration and improve the performance of the method for solving dynamic multi-objective optimization problems.(2)Source Domain Selection-Based Dynamic Multi-Objective Optimization Algorithm(STL-DMOEA).By quantifying the differences between multiple historical environmental data and the target domain,the one with the smallest difference is selected as the migration solution,the diversity solution is obtained by expanding the mutation rate,and the source domain data is constructed by nondominant sorting with the migration solution.The initial solution is set at the moment,thus a new learning framework based on source domain selection is proposed.The method is combined with the underlying non-dominated sorting genetic algorithm NSGA-Ⅱ,and the experimental results on 12 benchmark functions show that the proposed method can improve the performance to solve dynamic multi-objective optimization methods.