Reward Shaping based on Information Design in Multi-Agent Reinforcement Learning

The paper discusses the information-based reward shaping in multi-agent reinforcement learning. The architectures of reinforcement learning algorithms using various types of information are considered. The algorithm of multi-agent deep deterministic policy gradient using information-based reward shaping is presented. The architecture and key elements of the algorithm are shown: actor, critic and information function. The components of the algorithm and their interaction are described. The environment for training the algorithm is selected. Aspects of testing in the selected environment are shown. The results are demonstrated, according to which the algorithm success, the consequences of introducing the information function and the ability of the algorithm to determine the necessary policy to achieve the goal in the selected environment are determined. The work done is proof that the information-based reward shaping in multi-agent reinforcement learning allows to speed up and improve the algorithm results, as well as achieve the set goals in complex multi-agent environments.