MOEA/D with Adaptive Constraint Handling for Constrained Multi-objective Optimization

Most machine intelligence or cloud computing can be formulated as multi-objective optimization problems (MOPs) with constraints, while evolutionary multi-objective optimization (EMO) is a powerful means to deal with them. However, its adaptation for dealing with complex constrained MOPs (CMOPs) keeps being under the scope of recent investigations. The main challenges are as follows. 1) The existing algorithms can not make full use of infeasible solution information in the evolution process. 2) There is no effective infeasible solution in the initial population, which causes the algorithms fall into local optimal feasible regions. In light of these two issues, this paper proposes an adaptive epsilon-constraint-handling technique with a detect-and-escape strategy to make full use of infeasible solutions in the whole evolution process. Then, the feasible solutions are saved to the external archive and take part in the population evolution by non-dominated sorting. Finally, the proposed method is embedded into the decomposition based multi-objective evolutionary framework (MOEA/D). Experiments on benchmark problems show that the proposed algorithm is highly competitive compared with state-of-the-art constrained evolutionary algorithms.