A self-adaptive dynamic multi-objective optimization algorithm based on transfer learning and elitism-based mutation

Abstract

Dynamic multi-objective optimization problems (DMOPs) involve several conflicting objectives, and these objective functions change over time. Therefore, addressing DMOPs necessitates an effective response to environmental changes. However, most existing algorithms only deal with DMOPs with one particular type of environmental changes, whereas real-world dynamic changes are more complicated. Therefore, this paper proposes a self-adaptive DMOEA based on transfer learning and elitism-based mutation (ATM-DMOEA), aiming to efficiently tackle DMOPs exhibiting complex environmental changes. Specifically, a change evaluation method is devised to gauge change intensity and discern whether a change is drastic or gentle. Subsequently, an adaptive change response strategy is implemented to accommodate varying environmental changes. For drastic changes, the algorithm employs an elitism-based manifold transfer learning method, while gentle changes are handled with a diversity enhancement strategy introduced by adaptive elitism-based mutations with a varying mutation probability. The experiments have validated the competitiveness of the proposed ATM-DMOEA on the majority of DMOP test instances with different levels of change severity.