Multi-agent cooperation based on behavior prediction and reinforcement learning

In multi-agent systems based on reinforcement learning, the evaluation to the behavior of an agent depends on the other agents´ behaviors closely. The cooperation performance of multi-agent systems can be improved when each agent takes its action after it predicts the other agents´ actions self-consciously. In this paper, several methods for predicting other agents´ behaviors were presented, which demand all agents to evaluate the probabilities of actions that other agents may take, and joint-action was introduced to the traditional reinforcement learning. An experiment that three agents cooperate to raise an object was conducted to test the performance of multi-agent systems, and its results show that the cooperation process can be speeded by behavior prediction and joint-action is applied to the traditional reinforcement learning successfully.