A Multiobjective Evolutionary Multitasking Algorithm Based on Decomposition and Multiple Knowledge Transfer

Abstract

Multiobjective evolutionary multitasking (MOEMT) has become very popular in recent years, as this kind of methods aims to solve a set of multiobjective optimization problems (MOPs) simultaneously, which has been validated to be more promising than the traditional way that solves MOPs separately. However, most existing studies of MOEMT solve the MOP as a whole and use one single knowledge transfer strategy for solving all tasks, which is not so efficient as the MOP could be further decomposed into a set of subproblems and needs different strategies for knowledge transfer. To solve this problem, this paper suggests a new MOEMT algorithm (MOEMTA) based on decomposition and multiple knowledge transfer (MOEMTA-DM). First, the decomposition method is used to transform an MOP in each task into a set of subproblems, and the subproblems with greater performance improvement are chosen. Then their associated solutions will be selected to run multiple knowledge transfer strategies via both implicit and explicit ways, which can share the useful search experience between tasks. This way, computational resources can be adaptively assigned to speed up the solving of all tasks and the effect of knowledge transfer among all tasks can be improved. The effectiveness of the proposed algorithm was verified by studying benchmark multitasking MOPs (MTMOPs). Experimental results showed the proposed algorithm is more effective than other compared MOEMT algorithms.