Surrogate-Assisted Task Selection for Evolutionary Multitasking Optimization
Kaiyuan Huang, Xiaojun Wang, Yiqiao Cai · 2022
In the evolution algorithm, evolutionary multi-task optimization (EMTO) is recently proposed as a novel optimization paradigm for simultaneously solving multiple optimization tasks. However, in most of the existing EMTO algorithms, each offspring is evaluated selectively for only one task that is randomly assigned. This will lead to the negative transfer if the relatedness between tasks is lower. Based on this consideration, a novel surrogate-assisted task selection (SATS) method is proposed to assign a proper task to the offspring for evaluation by a collaborative multioutput Gaussian process surrogate (CoMOGP). The main purpose of SATS is to exploit the relatedness across tasks by jointly learning a surrogate model that is used for predicting the skill factor of each offspring, which will alleviate negative transfer. By incorporating the proposed SATS into a representative EMTO, multifactorial evolution algorithm (MFEA), an improved EMTO algorithm, termed SATS-EMO, is presented. Experimental results on a suite of single-objective multitasking benchmark problems have demonstrated the effectiveness of SATS for EMTO.