Multi-Agent Planning by Plan Set Intersection

Abstract

Coordination of a team of cooperative agents and their activities towards fulfillment of goals is described by multi-agent planning. For deterministic environments, where agents are not willing to share all their knowledge, the MA-STRIPS model provides minimal extension from classical planning. MA-STRIPS exactly prescribes what information can be freely communicated between the agents and what information has to be kept private such that the shared or individual goals can be still achieved. This thesis proposes a novel multi-agent planning approach which distributively intersects local plans of the agents towards a global solution of the multi-agent planning problem. This core principle builds on local compilation to a classical planning problem and compact representation of the local plans in the form of Finite State Machines. The efficiency of the resulting planner is further boosted up by distributed delete-relaxation heuristic, an approximative local plan analysis, and reduction of agents’ internal problems. The planning approach is analysed theoretically, in particular we prove its completeness and soundness. Experimental evaluation shows its applicability in a full privacy setting where only public information can be communicated and in less restricted privacy settings. At a recent international competition of distributed multiagent planners, the proposed planner showed top performance when compared with other existing state-of-the-art multi-agent planners.