Breaking Deadlocks in Multi-agent Reinforcement Learning with Sparse Interaction

Although multi-agent reinforcement learning (MARL) is a promising method for learning a collaborative action policy that will enable each agent to accomplish specific tasks, the state-action space increased exponentially. Coordinating Q-learning (CQ-learning) effectively reduces the state-action space by having each agent determine when it should consider the states of other agents on the basis of a comparison between the immediate rewards in a single-agent environment and those in a multi-agent environment. One way to improve the performance of CQ-learning is to have agents greedily select actions and switch between Q-value update equations in accordance with the state of each agent in the next step. Although this “GPCQ-learning” usually outperforms CQ-learning, a deadlock can occur if there is no difference in the immediate rewards between a single-agent environment and a multi-agent environment. A method has been developed to break such a deadlock by detecting its occurrence and augmenting the state of a deadlocked agent to include the state of the other agent. Evaluation of the method using pursuit games demonstrated that it improves the performance of GPCQ-learning.