Abstract
Intermittent sensory, actuation and communication failures may cause agents to fail in maintaining their commitments to others. Thus to collaborate robustly, agents must monitor others to detect coordination failures. Previous work on monitoring has focused mainly on small-scale systems, with only a limited number of agents. However, as the number of monitored agents is scaled up, two issues are raised that challenge previous work. First, agents become physically and logically disconnected from their peers, and thus their ability to monitor each other is reduced. Second, the number of possible coordination failures grows exponentially, with all potential interactions. Thus previous techniques that sift through all possible failure hypotheses cannot be used in large-scale teams. This paper tackles these challenges in the context of detecting disagreements among team-members, a monitoring task that is of particular importance to robust teamwork. First, we present new bounds on the number of agents that must be monitored in a team to guarantee disagreement detection. These bounds significantly reduce the connectivity requirements of the monitoring task in the distributed case. Second, we present YOYO, a highly scalable disagreement-detection algorithm which guarantees sound detection. YOYO's run-time scales linearly in the number of monitored agents, despite the exponential number of hypotheses. It compactly represents all valid hypotheses in single structure, while allowing for a complex hierarchical organizational structure to be considered in the monitoring. Both YOYO and the new bounds are explored analytically and empirically in monitoring problems involving thousands of agents.
Original language | English |
---|---|
Pages (from-to) | 501-525 |
Number of pages | 25 |
Journal | Autonomous Agents and Multi-Agent Systems |
Volume | 18 |
Issue number | 3 |
DOIs | |
State | Published - Jun 2009 |
Bibliographical note
Funding Information:Acknowledgments This work is based in part on a 2002 paper by the author and Michael Bowling [17]. We are indebted to Michael Bowling for his help in proving Theorem 2, for which he deserves joint credit. We also owe Michael many thanks for discussions and comments on earlier drafts of this work, including finding flaws with earlier attempted proofs. Milind Tambe and David V. Pynadath helped with initial versions of the YOYO algorithm. We also thank Meir Kalech, Dorit Avrahami-Zilberbrand, and Michael Lindner for useful discussions and corrections. As always, thanks to K. Ushi. This research was supported in part by Israel Science Foundation grants #1211/04, and #1357/07.
Funding
Acknowledgments This work is based in part on a 2002 paper by the author and Michael Bowling [17]. We are indebted to Michael Bowling for his help in proving Theorem 2, for which he deserves joint credit. We also owe Michael many thanks for discussions and comments on earlier drafts of this work, including finding flaws with earlier attempted proofs. Milind Tambe and David V. Pynadath helped with initial versions of the YOYO algorithm. We also thank Meir Kalech, Dorit Avrahami-Zilberbrand, and Michael Lindner for useful discussions and corrections. As always, thanks to K. Ushi. This research was supported in part by Israel Science Foundation grants #1211/04, and #1357/07.
Funders | Funder number |
---|---|
Israel Science Foundation | 1211/04, 1357/07 |
Keywords
- Collaboration
- Coordination
- Disagreement
- Exception handling
- Failure detection
- Failure handling
- Model-based diagnosis
- Observation-based coordination
- Plan recognition
- Robustness
- Teamwork