Reward-based participant selection for improving federated reinforcement learning

초록

Federated reinforcement learning (FRL) has recently received a lot of attention in various fields. In FRL systems, the concept of performing more proper actions with better experiences exists, and we focused on this unique characteristic. Motivated by such inherent property of FRL, in this paper, we propose the reward-based participant selection scheme for improving FRL. The FRL system with the proposed scheme performs learning effectively by putting a priority on utilizing better experiences of agents performing outstanding actions. We conducted various experiments, and the results show that it is possible to accelerate learning and require fewer agents when using the proposed scheme, which means that the proposed scheme improves the performance and efficiency of learning.(c) 2022 The Author. Published by Elsevier B.V. on behalf of The Korean Institute of Communications and Information Sciences. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).