动态惩罚分解策略下的高维目标进化算法

Many-objective Evolutionary Algorithm Based on the Dynamic Penalty Strategy

基于分解的多目标进化算法(MOEA/D)的求解精度与聚合方法有直接关系,其中基于惩罚的边界交叉聚合法(PBI)受惩罚参数θ影响较大,固定的惩罚参数难以匹配不同位置的子问题,尤其边界子问题上的极端解易被邻域内非支配解替代.为此,本文提出一种基于动态惩罚分解策略的高维目标进化算法(MOEA/D-DPS),通过动态调整惩罚参数θ来改变候选解选择区域大小,使不同位置的子问题都有更合适的选择区域,且减少了边界子问题上优秀解的丢失,更好地平衡了算法的收敛性与多样性.最后,本文通过仿真实验对比分析了MOEA/D-DPS算法与相关算法的性能,实验结果表明:MOEA/D-DPS算法在DTLZ1-4测试函数上所得解集整体性能更优.

The precision of the solutions generated by the multi-objective evolutionary algorithm based on decomposition （MOEA/D） is related to its decomposition approaches. The penalty-based boundary crossing strategy （PBI） of original MOEA/D is so much influenced by the penalty parameter θ that a fixed penalty parameter value is hard to befit the sub-problems all in different positions, and the extreme solutions situated in the boundary sub-problems are easily replaced by the non-dominated solutions in the neighborhood. In this paper, a many-objective evolutionary algorithm based on the dynamic penalty strategy （MOEA/D-DPS） is proposed, which alters the sizes of select area of the candidate solutions by dynamically adjusts the penalty parameter θ, provides the sub-problems in different positions with more appropriate select areas, reduces the loss of solutions with high quality that situated in the boundary sub- problems and keeps a better balance of convergence and diversity of the algorithm. Finally, this paper compares the performance of MOEA/D-DPS with related algorithms through simulation. The experimental results show that the MOEA/D-DPS algorithm is superior to the contrast algorithms in the DTLZ1-4 test functions.