Proposal of reward sharing method based on safety level and verification of its effectiveness in multi-agent environment
Naoki Kodama, Kazuteru Miyazaki, Hiroaki Kobayshi · 2017
In this paper, a reward sharing method based on safety level is proposed to improve the efficiency of the Profit Sharing based reinforcement learning method. To do so, a safety level is allotted to each rule, and the share of the reward for the rule is modified depending on the values. The share method of the relevant parameters is also shown. The efficiency is shown in multi-agent environment through a Keepaway task.