Transfer learning based evolutionary algorithm framework for multi-objective optimization problems

In this paper, a transfer learning based evolutionary algorithm (TLEA) framework for multi-objective optimization problems (MOPs) is proposed. In the TLEA framework, a complex multi-objective optimization task is decomposed into a set of relatively simple multi-objective optimization subtasks and then optimized collaboratively by parallel subpopulation searches with the proposed transfer learning method. More specifically, neighboring subtasks may have some similar features during parallel searches of corresponding subpopulations, and those similarities can be exploited through the proposed transfer learning strategy to improve the collaboration among these search subpopulations and achieve greater efficiency. To show the generality of the proposed algorithm framework, two implementations of the proposed TLEA framework based on differential evolution (DE) and particle swarm optimization (PSO), i.e., TLPSO and TLDE, are presented and studied in detail. In TLPSO and TLDE, the subproblem features are reflected by the search subpopulations, which are generated by a pair of specific parameters. Therefore, subpopulations can adaptively adjust parameter settings by learning useful information from neighboring subproblems with more appropriate parameters during the search. The experimental results show that TLPSO performs better than other algorithms on at least five out of 12 test problems in terms of the IGD indicator and on at least seven out of 12 test problems in terms of the HV indicator. TLDE has an advantage over the other algorithms on five out of 12 test problems in terms of the IGD indicator and on seven out of 12 test problems in terms of the HV indicator.