Research On Multi-objective Multi-task Evolutionary Optimization Algorithm Based On Feature Transfer

In practice,evolutionary algorithms are often applied to various engineering problems.Traditional evolutionary algorithms tend to deal with problems sequentially,so each problem needs to be optimized from scratch.However,many optimization problems are similar,and the knowledge gained from solving one problem can help solve other similar problems.In order to improve computational efficiency,multi-task evolutionary algorithm based on parallel processing can be adopted to promote knowledge transfer among similar problems.In order to improve the efficiency of knowledge exchange between tasks,this paper conducts in-depth research on knowledge transfer,and the main research contents are as follows:(1)To solve the problem of negative knowledge transfer between tasks,a multiobjective multi-task optimization algorithm(IM-MFEA)based on inverse model mapping and objective space alignment strategy was proposed.Firstly,the objective space alignment strategy is used to improve the quality of source domain solutions in the objective space.The source domain solutions of these transformations are reconstructed by inverse mapping strategy.The reconstructed source domain solution is used to assist the target domain to generate competitive offspring.In order to verify the effectiveness of the algorithm,a comprehensive experiment is carried out on nine multi-objective multi-task benchmark problems.Experimental results show that this algorithm can suppress the probability of negative knowledge transfer to a large extent.According to IGD value index,IM-MFEA algorithm outperforms other multi-task algorithms in 90% of test cases.(2)In order to accurately quantify the similarity between tasks,a multi-task algorithm(MTO-SD)is proposed to calculate the similarity between tasks in real time.The algorithm uses a two-space similarity calculation strategy to calculate the global similarity between tasks in real time.In order to deal with different similarity,the algorithm designs two knowledge transfer strategies suitable for different similarity.The algorithm will choose the appropriate knowledge exchange strategy according to the similarity.In order to cope with different situations of the population,IOT indicators are designed for adaptive selection of suitable DE operators to complete mating and achieve evolution.In order to verify the effectiveness of the algorithm,the algorithm is tested on the standard multi-task test function.Experimental results show that the algorithm has super convergence ability and extremely fast convergence speed.