[4G1-GS-2j] 機械学習：要素技術

[4G1-GS-2j-02] Curiosity-Driven Search in a Multiagent Reinforcement Learning Problem

〇Toru Iwashina¹, Koichi Moriyama¹, Tohgoroh Matsui², Atsuko Mutoh¹, Nobuhiro Inuzuka ¹ (1. Department of Engineering, Graduate School of Engineering, Nagoya Institute of Technology, 2. Department of Clinical Engineering, College of Life and Health Sciences, Chubu University)

Keywords:Reinforcement Learning

Reinforcement learning has been shown to be capable of dealing with complex control problems in the real world, such as automated driving. On the other hand, in order for individual agents to learn in an environment with multiple agents, such as the real world, efficient search in the strategy space, which increases with the number of agents, is necessary. In this study, we attempted to improve the learning efficiency by introducing curiosity search, which is an efficient search method for reinforcement learning of a single agent, to reinforcement learning in a multi-agent environment. We conducted experiments using a tracking problem, which is a typical problem in a multi-agent environment, and found that the learning speed in the early stage of learning was improved compared to the case without the introduction of curiosity search.