When and How to Transfer Knowledge in Dynamic Multi-objective Optimization

Transfer learning has been used for solving multiple optimization and dynamic multi-objective optimization problems, since transfer learning is able to transfer useful information from one problem to help solving another related problem. This paper aims to investigate when and how transfer learning works or fails in dynamic multi-objective optimization. Through computational analyses on a number of dynamic bi- and tri-objective benchmark problems, we show that transfer learning fails on problems with fixed Pareto optimal solution sets and under small environmental changes. We also show that the Gaussian kernel function used in the existing transfer learning-based method is not always adequate. Therefore, transfer learning should be avoided when dealing with problems for which transfer learning fails and other kernel functions should be used when the Gaussian kernel is inadequate. This paper proposes novel strategies and kernel functions that can be used in such cases. Experimental studies have demonstrated the superiority of our proposed techniques to state-of-the-art methods, on a number of dynamic bi- and tri-objective test problems.