A Dynamic Multi-objective Evolutionary Algorithm Assisted by Kernel Ridge Regression

Affected by uncertainty environments, the objective functions and constraints or the associated problem parameters of the dynamic multi-objective optimization problems (DMOPs) in the real world may change over time. Meanwhile, the Pareto optimal solution set of DMOPs is also often time-varying. In this paper, a kernel ridge regression assisted multi-objective evolutionary algorithm based on decomposition (MOEA/D-KRR) is proposed to effectively solve a kind of DMOPs with variable linkage and nonlinear characteristics. Firstly, a new prediction strategy based on kernel ridge regression is designed to help the proposed algorithm adapt to the new change quickly. The new prediction strategy not only uses ridge regression to reduce adverse effects of multicollinearity caused by variable linkage, but also adopts a Gaussian kernel function to improve nonlinear processing capability of the proposed algorithm. Next, the new prediction strategy is integrated into a multi-objective evolutionary algorithm based on decomposition with a differential evolution operator to handle DMOPs. Finally, the proposed MOEA/D-KRR is compared with other three dynamic multi-objective evolutionary algorithms assisted by state-of-the-art prediction strategies on 12 benchmark problems. The experimental results show that MOEA/D-KRR is promising for solving nonlinear DMOPs with variable linkage.