Reinforcement learning aided parameter control in multi-objective evolutionary algorithm based on decomposition

• Published in: Progress in artificial intelligence Vol. 7; no. 4; pp. 385 - 398
• Main Authors: Ning, Weikang, Guo, Baolong, Guo, Xinxing, Li, Cheng, Yan, Yunyi
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.12.2018; Springer Nature B.V
• Subjects: Artificial Intelligence; Computational Intelligence; Computer Imaging; Computer Science; Control; Data Mining and Knowledge Discovery; Decomposition; Evolutionary algorithms; Genetic algorithms; Machine learning; Mechatronics; Multiple objective analysis; Natural Language Processing (NLP); Parameters; Pattern Recognition and Graphics; Regular Paper; Robotics; Vision; Multi-objective optimization; Decomposition; Parameter control; Reinforcement learning
• ISSN: 2192-6352; 2192-6360
• DOI: 10.1007/s13748-018-0155-7

Multi-objective evolutionary algorithm based on decomposition (MOEA/D) has been successfully applied in solving multi-objective optimization problems. However, the performance of MOEA/D could be severely influenced by its parameter settings. In this paper, we introduce reinforcement learning into MOEA/D as a generic parameter controller. The resulting algorithm, reinforcement learning enhanced MOEA/D (RL-MOEA/D), is used to adaptively control the neighborhood size T and the differential evolutionary operators used in MOEA/D. RL-MOEA/D is first compared with MOEA/D with a random parameter control mechanism and MOEA/Ds with some fixed parameter settings on ten widely used multi-objective test instances. Then, RL-MOEA/D is compared with FRRMAB to show the effectiveness of the proposed algorithm. The experimental results indicate that RL-MOEA/D is very competitive. Finally, the characteristics of RL-MOEA/D are studied.