Broadening the Scope of Multi-Agent Plan Recognition: Theory and Practice
Summary
Plan Recognition is the problem of inferring the goals and plans of an agent given a set of observations. In Multi-Agent Plan Recognition (MAPR) the task is extended to inferring the goals and plans of multiple agents. Previous MAPR approaches have made various strong assumptions which have limited their applicability to a restricted set of real-world instantiations of the MAPR problem. In order to broaden the applicability of MAPR to a wider range of problems, in this thesis we characterize two novel formulations of the MAPR problem, each relaxing different assumptions made by previous work. The first formulation defines the Epistemic MAPR problem, which no longer assumes that all agents must share a common mental state. This, in turn, enables the observing agent to consider the unique perspective of each observed agent when detecting its likely plans and goals. The second formulation defines the MAPR problem with temporal actions and unreliable observations. This formulation relaxes the assumptions that (a) the agents' actions are instantaneous and (b) the observations are perfect and reliable. Importantly, the thesis proposes to conceive the computational core of the MAPR problem as an AI planning task, thus enabling the use of existing planning tools. The thesis then introduces different AI planning-based computational approaches which solve the novel formulations of the MAPR problem by solving the corresponding planning problems. Finally, the thesis illustrates the power and flexibility of the proposed computational approaches by demonstrating their applicability to various, previously unaddressed, MAPR problems.