Greedy Action Selection and Pessimistic Q-Value Updates in Cooperative Q-Iearning
Toshihiro Kujirai, Takayoshi Yokota · 2018
Multi-agent reinforcement learning (MARL) is a promising method that can deal with learning collaborative action policy, enabling each agent to accomplish tasks. In MARL, an assumption of sparse interactions dramatically reduces the state/action space for learning. This paper proposes an improvement in CQ-Iearning, which is a typical method for sparse interaction tasks, taken from the perspective of action selection and updating augmented Q-values. On the basis of the results of several maze games, we concluded that the performance can be improved using greedy action selection and switching Q-values updating equations based on the state of each agent in the next step.