Cooperative Action Acquisition Based on Intention Estimation in a Multi-Agent Reinforcement Learning System

In this paper, we propose a method to acquire a series of cooperative actions to reach an appropriate goal without the designer controlling the reward. To accomplish this, we introduce a new concept of "reward interpretation." This is the idea that an agent can increase or decrease the reward given by the environment through the reward interpretation on its won. We applied this idea to the Q-learning method. The simulation results show that the proposed method is superior to a standard Q-learning method and a Q-learning method with cooperation in terms of the number of successful instances of cooperation. (C) 2017 Wiley Periodicals, Inc.