Dynamic agent-based reward shaping for multi-agent systems

Earlier works have reported that reward shaping accelerates the convergence of reinforcement learning algorithms. It also helps to make better use of existing information. In this article we propose the use to modify Q-learning in multiagent systems by the use of reward shaping depending on agent state regarding other agents. We study this method with different choices, which indicate different effects of this method on the maze problem. The results indicate the directional search, reduces the number of steps to reach the target in the proposed modified approach if appropriate parameters are utilized.