Discrete Interactions in Decentralized Multiagent Coordination: A Probabilistic Perspective

While current work has proved that the interactions between tightly coupled multiagent is one of the effective means to overcome the information partially observable constraints, creating coordinative policies among loose coupled agents under uncertainty is still a big challenge. In this article, we explored how the discrete interactions affect the decentralized multiagent coordination, in which the coordinative relationship changes dynamically with randomly arrived tasks. We first proposed a decision model, DDI-MDPs, which supports both the independence and the interactive decision making, that is, interactions are spatiotemporal discrete occurred. On this basis, we contributed a heuristic imprecise probabilistic-based interaction algorithm, HDLI, which utilizes the multidimensional semantic relevance among agents, information and tasks, so that agents can continuously improve their cognition and optimize the decision-making efficiency by learning the interaction records. In addition, we dissected the transformation conditions between DDI-MDPs and the classic models, and modeled the dynamic relationships into a dynamic graph, and then analyzed its evolution processes. In the simulation, we evaluated the efficiency of the proposed work in several typical coordination scenarios, the results reveal that the interactive-growth relationship between agents has typical social network characteristics. Finally, some possible challenges in the follow-up research work and applications are discussed.