In many multi-agent domains, the interaction amongintelligent agents -- collaborative or non-collaborative-- is both dynamic and real-time. Examples includeintelligent tutoring systems that interact with studentsin real-time(Anderson et al. 1990), virtual environ-ments for entertainment(Maes et ai. 1994; Bates, Loy-all, & Reilly 1992), "real-world" synthetic environ-ments for training in traffic(Cremer et al. 1994) orbattlefield simultations(Tambe et al. 1995), RoboCupsoccer(Kitano et al. 1995), and robotic collabora-tion by observation(Kuniyoshi et al. 1994). In mostsuch domains, agent modeling is crucial for effectivecollaboration and competition. The key aspect ofagent modeling of interest here is agent tracking --monitoring other agents’ observable actions and infer-ring their high-level goals, plans and behaviors(An-derson et al. 1990; Tambe & Rosenbloom 1995;Rao 1994). In contrast with plan recognition in morestatic domains(Kautz & Allen 1986), this capabilityfocuses on tracking flexible/reactive behaviors in dy-namic environments.Previous work in agent tracking has mostly focusedon tracking individual agents. This paper outlines keychallenges that arise in going beyond individuals totracking agent teams; and presents some initial so-lutions to adress those challenges 1. Already, a largenumber of multi-agent systems, both in synthetic androbotic domains, require that agents understand andtrack team activities (e.g., many of the domains de-scribed above, including RoboCup). The basic chal-lenge in tracking teamwork is that it is more thana group of individual agents acting simultaneously,even if in a coordinated fashion(Grosz & Sidner 1990;Cohen & Levesque 1991; Kinny et al. 1992; Jennings1995). For instance, driving in ordinary traffic is notconsidered teamwork, despite the drivers’ simultaneousactivity, coordinated by traffic signs(Cohen & Levesque1991). Conversely, team activity may not be decom-posed and tracked as independent actions of individu-als. Thus, for instance, if two children are collabora-1 Portions of this paper are based on (Tambe 1996b).tively building a tower of blocks(Grosz & Sidner 1990),then they cannot be tracked as building two individ-ual towers of blocks with gaps in just the right places.Similarly, a wall pass in Soccer cannot be tracked asindividuals’ independent actions -- no one individualexecutes the wall pass (for non soccer literates, thenext section provides an explanation).While researchers broadly agree that team activityis not merely coordinated activity, its precise natureis still a topic of active research and debate(Jennings1995). Nonetheless, this paper will rely on one lead-ing theory of teamwork that is based on joint inten-tions, i.e., joint commitments to joint activities in ashared belief state(Cohen & Levesque 1991). Jointcommitment implies that team members have a mu-tual belief that they are each committed to that ac-tivity. Furthermore, a team’s jointly intending an ac-tivity leads subteams to intend to do their share inthat activity, subject to the joint intention remainingvalid. Finally, should a team member privately dis-cover that the team’s joint activity is achieved, un-achievable or irrelevant, it must inform other teammembers -- so this private knowledge becomes mu-tual knowledge. This communication is an overhead ofteamwork, which team members bear so as to save theteam’s time/resources(Jennings 1995).While the joint intentions framework attempts tooutline ideal behavior for team members, its specifiedcommunication requirement can be problematic if suchcommunication is costly, risky or redundant. In Soc-cer, for instance, a player usually may not communi-cate the failure to execute a pass to its teammate --he/she may waste an opportunity to score a goal in theprocess. Thus, as ideal teammembers, agents shouldbalance communication costs and benefits (this is anextension to the joint intentions framework). We willrefer to teams that adhere to the extended joint inten-tions framework as pure teams.2 In pure teams, agentsfully adhere to the jointness of the team, and engage2One interesting issue here is whether this definition issufficient to establish the pureness of a team -- specifically,team members do not appear to share risks, resources andrewards.83
[1]
A. Newell.
Unified Theories of Cognition
,
1990
.
[2]
Nicholas R. Jennings,et al.
Controlling Cooperative Problem Solving in Industrial Multi-Agent Systems Using Joint Intentions
,
1995,
Artif. Intell..
[3]
Milind Tambe,et al.
Building Intelligent Pilots for Simulated Rotary Wing Aircraft
,
1995
.
[4]
A. B. Loyall,et al.
Integrating Reactivity, Goals, and Emotion in a Broad Agent
,
1992
.
[5]
Henry A. Kautz,et al.
Generalized Plan Recognition
,
1986,
AAAI.
[6]
John R. Anderson,et al.
Cognitive Modeling and Intelligent Tutoring
,
1990,
Artif. Intell..
[7]
Milind Tambe,et al.
Architectures for Agents that Track Other Agents in Multi-Agent Worlds
,
1995,
ATAL.
[8]
Milind Tambe,et al.
Tracking Dynamic Team Activity
,
1996,
AAAI/IAAI, Vol. 1.
[9]
A. B. Loyall,et al.
Integrating Reactivity , Goals , and Emotionin a Broad AgentJoseph
,
1992
.
[10]
Masaru Ishii,et al.
Cooperation by observation: the framework and basic task patterns
,
1994,
Proceedings of the 1994 IEEE International Conference on Robotics and Automation.
[11]
Milind Tambe.
Executing Team Plans in Dynamic, Multi-agent Domains
,
1996
.
[12]
Anand S. Rao,et al.
Means-End Plan Recognition - Towards a Theory of Reactive Recognition
,
1994,
KR.
[13]
Milind Tambe,et al.
Intelligent Agents for Interactive Simulation Environments
,
1995,
AI Mag..
[14]
Allen Newell,et al.
SOAR: An Architecture for General Intelligence
,
1987,
Artif. Intell..
[15]
James F. Cremer,et al.
The Software Architecture for Scenario Control in the Iowa Driving Simulator
,
1993
.
[16]
Hiroaki Kitano,et al.
RoboCup: The Robot World Cup Initiative
,
1997,
AGENTS '97.
[17]
C. Sidner,et al.
Plans for Discourse
,
1988
.
[18]
Milind Tambe,et al.
RESC: An Approach for Real-time, Dynamic Agent Tracking
,
1995,
IJCAI.
[19]
Gil Tidhar,et al.
Planned Team Activity
,
1992,
MAAMAW.
[20]
M. F. Mar,et al.
ModSAF Behavior Simulation and Control
,
1993
.