Evolutionary Multitasking via Reinforcement Learning
Shuijia Li, Wenyin Gong, Ling Wang, Qiong Gu
Abstract
Different from traditional evolutionary algorithms (EAs), the multifactorial evolutionary algorithm (MFEA) is proposed to optimize multiple optimization tasks concurrently. Through the knowledge transfer between different tasks, MFEA has been proved to be superior to single-task EAs in the solution quality and convergence speed. Recently, various MFEAs have been developed. Most of them are based on a common model in MFEA, where a fixed knowledge transfer parameter, i.e., random mating probability ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$rmp$</tex-math></inline-formula> ), is used. In addition, a single evolutionary search operator is employed in the whole evolutionary process. However, in this model, the fixed <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$rmp$</tex-math></inline-formula> is difficult to adapt to multiple different tasks. Besides, a single evolutionary search operator may not be suitable for problems with different properties, thus limiting the performance of the algorithm. Based on these considerations, in this article, a reinforcement learning based multifactorial evolutionary algorithm (RLMFEA) is presented. In RLMFEA, it allows different evolutionary search operators to be embedded in MFEA, and each task has a changing <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$rmp$</tex-math></inline-formula> that is adaptively adjusted by reinforcement learning. The effectiveness of RLMFEA has been verified on a series of single-objective multitask optimization benchmark functions and a real-world application.