Exploring Information Interactions in Decentralized Multiagent Coordination under Uncertainty

Information interactions between tightly coupled agents has been proved to be an effective method to overshadow the partially observable constraints and simplify the solving complexity. In this work, we explore how incompact information interaction affect the decentralized multiagent coordination under uncertainty. We focus on problems in which the interactions between agents are spatio-temporal discrete, and propose a practical decision model for decentralized multiagent coordination, DLI-MDPs, that both supports the decision-making under the state of none and restricted information interactions. We analysis the relevances and transformation conditions between the proposed new model and the classical models, then contribute a reinforcement learning algorithm HDLI, that takes the advantage of the particular structure of DLI-MDPs in which agent updates policies by learning both its local cognition and the additional information obtained through interaction. Finally, we compare and verify the application of the algorithm throughout the paper in several typical coordination scenarios.