Multi-Agent Reinforcement Learning Based on Bidding

In a multi-agent environment, if multiple agents learn simultaneity, the feedbacks of the environment would be confusing, even be conflicting. This paper presents an approach for developing multi-agent reinforcement learning systems in which all agents learn alternately. In each learning cycle, only the active agent executes the action calculated by the reinforcement learning algorithm and is in the state of learning phase. All the other agents take actions acquired previously and are in the state of non-learning phase. After the active agent finishes the learning phase, another agent is chosen to learn by bidding. The proposed method has been implemented in soccer game and the high efficiency of the proposed scheme was verified by the result of computer simulation.