Monitoring teams by overhearing: a multi-agent plan-recognition approach

摘要: Recent years are seeing an increasing need for on-line monitoring of teams cooperating agents, e.g., visualization, or performance tracking. However, in deployed teams, we often cannot rely on the agents to always communicate their state system. This paper presents a non-intrusive approach by overhearing, where monitored team's is inferred (via plan-recognition) from team-members' routine communications, exchanged as part coordinated task execution, and observed (overheard) Key challenges this include demanding run-time requirements monitoring, scarceness observations (increasing uncertainty), scale-up address potentially large teams. To these, present set complementary novel techniques, exploiting knowledge social structures procedures team: (i) efficient probabilistic plan-recognition algorithm, well-suited processing communications observations; (ii) behavior predict future during execution (reducing uncertainty); (iii) algorithms that trade expressivity scalability, representing only certain useful hypotheses, but allowing any number different activities be represented single coherent entity. We empirical evaluation these combination apart, team running machines physically distributed across country, engaged complex, dynamic execution. also compare techniques human expert novice monitors, show presented capable at human-expert levels, despite difficulty task.