Multiagent systems: games and learning from structures

Multiple agents have become increasingly utilized in various fields for both physical robots and software agents, such as search and rescue robots, automated driving, auctions and electronic commerce agents, and so on. In multiagent domains, agents interact and coadapt with other agents. Each agent's choice of policy depends on the others' joint policy to achieve the best available performance. During this process, the environment evolves and is no longer stationary, where each agent adapts to proceed towards its target. Each micro-level step in time may present a different learning problem which needs to be addressed. However, in this non-stationary environment, a holistic phenomenon forms along with the rational strategies of all players; we define this phenomenon as structural properties.

In our research, we present the importance of analyzing the structural properties, and how to extract the structural properties in multiagent environments. According to the agents' objectives, a multiagent environment can be classified as self-interested, cooperative, or competitive. We examine the structure from these three general multiagent environments: self-interested random graphical game playing, distributed cooperative team playing, and competitive group survival. In each scenario, we analyze the structure in each environmental setting, and demonstrate the structure learned as a comprehensive representation: structure of players' action influence, structure of constraints in teamwork communication, and structure of inter-connections among strategies. This structure represents macro-level knowledge arising in a multiagent system, and provides critical, holistic information for each problem domain. Last, we present some open issues and point toward future research.