A domain adaptation learning strategy for dynamic multiobjective optimization

Dynamic multiobjective optimization problems (DMOPs) require the robust tracking of Pareto-optima varying over time. Previous transfer learning-based problem solvers consume the most time on complex training of transfer model or applying a plenty of evaluations to find transferred individuals, decreasing computational efficiency. To address this issue, a domain adaptation learning strategy based dynamic multiobjective evolutionary algorithm is proposed in this paper. The mapping matrix learned by subspace distribution alignment (SDA) is utilized to transform the search space between last and current environments for promoting efficient knowledge transfer. Especially, the process of constructing mapping is derived from the simpler calculation, saving computational cost. Based on this model, transferred individuals are generated from a part of historical optima at last time. Additionally, an increment information is defined as the difference between center points of POSs in past two environments, and employed to produce a noise obeying uniform distribution. After adding it on a temporary population consisting of transferred individuals and the rest historical optima, an initial population with good diversity under new environment is formed. Experimental results on 12 benchmark functions indicate that the proposed method outperforms the other six state-of-the-art comparative ones, achieving the promising performance in solving DMOPs.