M. Tambe. Tracking dynamic team activity. In Proc. AAAI, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....is formal and works for monitoring arbitrary programs. While we focus on monitoring the collaboration of multiple agents, they address the problem of a single agent acting in an uncertain environment. Another interesting monitoring approach is based on multi agent plan recognition, by Tambe [18], Intille and Bobick [12] Devaney and Ram [1] Kaminka et al. 13, 14] In this approach, an agent s intentions (goals and plans) beliefs or future actions are inferred through observations of another agent s ongoing behavior. Devaney and Ram [1] describe the plan recognition problem in a ....

....on all coordination constraints among the agents. Once an agent fails, it may not be able to recognize the plans. Another line of work has been pursued in ISI. Gal Kaminka et al. 13, 14] developed the OVERSEER monitoring system, which builds upon work on multi agent planrecognition by [12] and [18]. They address the problem of many geographically distributed team members collaborating in a dynamic environment. The system employs plan recognition to infer the current state of agents based on the observed messages exchanged between them. The basic component is a probabilistic plan recognition ....

M. Tambe. Tracking dynamic team activity. In Proc. AAAI-96, pp. 80--87, 1996. 15

....these conditions require the coordinating agent to not affect the behavior of the observed agents a significant restriction in collaborative and adversarial settings. Tambe developed the non probabilistic polynomial time RESC team algorithm for reasoning about adversaries hierarchical behaviors [10]. This algorithm gains its computational advantage by always adapting and committing to a single interpretation of the opponents actions. Independently from investigations of SBC (mostly in software agent settings) researchers in the multi robot community have focused their efforts on ....

Milind Tambe. Tracking dynamic team activity. In Proceedings of the National Conference on Artificial Intelligence (AAAI), August 1996.

....conditions require the coordinating agent to not affect the behavior of the observed agents a significant restriction in collaborative and adversarial settings. Tambe developed the non probabilistic polynomial time 99 9 algorithm for reasoning about adversaries hierarchical behaviors [10]. This algorithm gains its computational advantage by always adapting and committing to a single interpretation of the opponents actions. Independently from investigations of SBC (mostly in software agent settings) researchers in the multi robot community have focused their efforts on ....

Milind Tambe. Tracking dynamic team activity. In Proceedings of the National Conference on Artificial Intelligence (AAAI), August 1996.

....by B, P 2 M(A;B=P ) If A is monitoring a team of agents B 1 ; Bn , we say that A s team monitoring of the team is complete if A s monitoring of each of B 1 ; Bn is complete. Monitoring completeness is commonly assumed (in its individual form) in planrecognition work, e.g. [19, 6, 11]) and generally holds in our own applications. It means that the set M(A;B=P ) includes the correct hypothesis P , but will typically include other matching hypotheses besides P . Using this notation, we can now formally explore the role of key agents in disagreement detection. Key agents have ....

....tasks on which YOYO performs well; and (iii) YOYO includes explicit checks for failure detection. Also, while YOYO has revealed a potential tradeoff between expressivity and scalability in visualization algorithms, YOYO provides the same detection results as the full array approach. RESC team [19] is a multi agent plan recognition scheme which implicitly uses coherence as a key constraint in representation. RESC team represents only a single coherent hypotheses, while YOYO represents all coherent hypotheses. However, RESC team can reason about the assignment of agents to roles subteams, ....

Milind Tambe. Tracking dynamic team activity. In Proceedings of the National Conference on Artificial Intelligence (AAAI), August 1996.

....communication systems. An extended review is outside the objectives of this paper, however it is worthwhile to give some examples just to explain how channeled multicast can be bene cial. A rst example is monitoring the activity of other agents, or even of teams of agents; see, for instance, [14, 19]. Monitoring is critical, for example, for visualization [17] identi cation of failures, or simply to trace teams activity. One possible approach (report based monitoring) requires each monitored agent to explicity communicate its state to the monitoring system. Clearly, if reporting is done by ....

M. Tambe. Tracking dynamic team activity. In Int. Conf. on Articial Intelligence (AAAI96), 1996.

....Veloso [73] Schneider et al. 65] Wei [83] and Brauer and Wei [9] Others have pursued methods of coalition formation to optimize multiagent coordination. Examples are work done by Sandholm and Lesser [58] Sandholm et al. 60] Zlotkin and Rosenschein [89] and Shehory and Kraus [68] Tambe [77] proposes tracking behavior of other agents as an optimization technique. Rosin and Belew [55] and Matos and Sierra [47] adopt evolutionary methods for enhancing performance. Other optimization techniques have been proposed by Lux and Marchesi [41] Huber and Durfee [32] Castelfranchi and Conte ....

Tambe, M., "Tracking Dynamic Team Activity", Proceedings of the Thirteenth National Conference on Artificial Intelligence, 1996.

....relevant capabilities. In contrast, Murdoch [7] completely distributes the matchmaking process and thus does away with the centralized broker. A di erent approach is taken in STEAM [8] this modelbased system relies on each agent s explicitly tracking the actions of both itself and its team [9] and then applying joint intention principles to the resulting world model in order to make decisions. While this approach has been validated on software agents, it is not clear how, in a physical environment, an autonomous agent could gather and interpret the necessary data to perform the ....

Milind Tambe, \Tracking dynamic team activity," in Proceedings of the Natl. Conf. on Articial Intelligence (AAAI), Portland, Oregon, July 1996.

....for training are unavailable. RECOGNIZING PLANNED, MULTIPERSON ACTION 421 FIG. 4. Three play diagrams for examples used in this work: a) p52maxpin, b) p51curl, and (c) p56yunder. Although some logical backtracking search systems for recognizing multiagent goals and actions have been proposed [3, 42, 46], noisy visual data require a representation that can handle uncertainty. Pairwise comparison of features between trajectories can be used to recognize some group military behaviors for large numbers of agents [9] Huber has shown that simple goal recognition belief networks can be constructed ....

M. Tambe, Tracking dynamic team activity, in Proc. Nat. Conf. on Artificial Intelligence, pp. 80--87, August 1996.

....in our case) is added to every model and then the distribution is renormalized. This prevents any model s probability from going to 0, while not changing which model is most likely on any one update. 6 Related Work A great deal of work has been done in the area of plan recognition. For example, Tambe[1996] has explored tracking the high level intentions of agent teams. This can be useful to infer events which are not directly observable by an agent. However, his work requires knowledge of the workings of the other agent through operator hierarchies. We would like to be able to relax that ....

Milind Tambe. Tracking dynamic team activity. In AAAI-96. AAAI Press, 1996.

....used is W (p; w) pw (1 w) which implements what is shown in Figure 3. In doing the model probability updating, we apply the W function to each P [e j jS; B; M i ] term before normalizing. 6. RELATED WORK A great deal of work has been done in the area of plan recognition. For example, Tambe[18] has explored tracking the high level intentions of agent teams. This can be useful to infer events which are not directly observable by an agent. However, his work requires knowledge of the workings of the other agent through operator hierarchies. We would like to be able to relax that ....

M. Tambe. Tracking dynamic team activity. In Proceedings of the National Conference on Aritical Intelligence (AAAI-96). AAAI Press, 1996.

....a multi object scene [9] We build on this work. None of the proposed representations can easily represent fuzzy temporal relationships among multiple goals of multiple agents. Search based systems designed specifically to recognize multi agent goals and actions outside of probabilistic frameworks [17, 2, 18] are sensitive noisy data and detectors. We are aware of no prior systems that richly represent time and multi agent interaction in a probabilistic framework that can perform recognition of complex action on noisy, perceptual input data. 6 Final remarks We have proposed a representation ....

M. Tambe. Tracking dynamic team activity. In Proc. Nat. Conf. on Artificial Intelligence, pages 80--87, August 1996.

....and secure communications between the team members and the monitoring agent. Unfortunately, this is often not possible. Failures do occur, and communications must often be limited in adversarial settings for security reasons. An alternative monitoring approach is based on plan recognition (e.g. [17, 8, 11]) The monitoring agent infers the unobservable state of the agents based on their observable actions, using knowledge of the plans that give rise to the actions. This approach is completely non intrusive, requiring no changes to agents behaviors, and allows for changes in the requested ....

....that the techniques presented result in a significant boost to OVERSEER s monitoring accuracy and efficiency, while sacrificing little in terms of being able to detect failures. While previous work in multi agent plan recognition has either focused on exploiting explicit teamwork reasoning, e.g. [17], or explicitly reasoning about uncertainty, e.g. 8] a key novelty in OVERSEER is that it effectively blends these two threads together. 2. AN EXAMPLE DISTRIBUTED TEAM OVERSEER has been applied in multiple applications, monitoring distributed teams of heterogeneous, software agents, ....

[Article contains additional citation context not shown here]

Milind Tambe. Tracking dynamic team activity. In Proceedings of the National Conference on Artificial Intelligence (AAAI), August 1996.

....5 yards P2 QB C P4 P1 P3 RG (a) b) c) FIG. 4. Three New England Patriots play diagrams for examples used in this work: a) p52maxpin, b) p51curl, and (c) p56yunder. Although some logical backtracking search systems for recognizing multi agent goals and actions have been proposed [42, 3, 46], noisy visual data requires a representation that can handle uncertainty. Pairwise comparison of features between trajectories can be used to recognize some group military behaviors for large numbers of agents [9] Huber has shown that simple goal recognition belief networks can be constructed ....

M. Tambe. Tracking dynamic team activity. In Proc. Nat. Conf. on Artificial Intelligence, pages 80--87, August 1996.

....network training procedure for Dribble vs. Pass Team Strategy Layer This layer of learning investigates the performance of emergent team behavior. In order to adapt to different opponents and different scenarios, an effective team must learn to become more defensive or offensive during the match (Tambe 1996). In this particular scenario, a modelbased approach to team strategy is adopted (Tambe 1996, Tambe et al. 1998) as opposed to a strictly behavior based strategy (Werger 1999) This becomes the first layer of cooperation between the agents, and can potentially be expanded to further layers. Three ....

....investigates the performance of emergent team behavior. In order to adapt to different opponents and different scenarios, an effective team must learn to become more defensive or offensive during the match (Tambe 1996) In this particular scenario, a modelbased approach to team strategy is adopted (Tambe 1996, Tambe et al. 1998) as opposed to a strictly behavior based strategy (Werger 1999) This becomes the first layer of cooperation between the agents, and can potentially be expanded to further layers. Three types of strategies; offensive, defensive, and half half, are defined by three possible ....

Tambe, M. 1996. Tracking Dynamic Team Activity. In Proceedings of the thirteenth Conference on Artificial Intelligence Applications. 11:80-87. Cambridge, U.S.A.: MIT Press.

....believed. 3 Hence, the need for some form of communication is implicit in this model. 4 The form of this communication is domain dependent; in the case of a soccer team, communication could be either verbal or hand waving, using some prearranged signals. There exist other domains (see, e.g. [26, 27]) where, communication, while still vital to a team s success, has to be weighed against other factors, thereby revising the absolute requirement to communicate. The bottom line is that in all but the most unusual circumstances should an agent recognize his responsibility to make his private ....

.... of what could go wrong when an agent observes, draws inferences, and acts (incorrectly) instead of communicating directly is given as a convoy example in [5] This theory would be valuable in a domain where communication is either unreliable, or where it must be abandoned altogether (see, e.g. [26, 27]) Matsubayashi [20] explores the relationship between two agents, one of whom wishes to delegate his responsibilities to another. This delegation arises because in the development of an individual agent s own actions and goals, there is bound to be overlap with another agent s if both are part ....

[Article contains additional citation context not shown here]

Tambe, M. "Tracking Dynamic Team Activity", Proc. Natl. Conf. Artificial Intelligence (AAAI), 1996.

....probabilistic plan recognition have not been demonstrated on problems where it is necessary to represent fuzzy temporal relationships among multiple goals of multiple agents. Search based systems designed specifically to recognize multi agent goals and actions outside of probabilistic frameworks [28, 3, 31] are sensitive to noisy data and detectors. Belief networks have been used for visual recognition of static objects [20, 1] and for visual attention selection[29] Promising work on recognizing single agent action from trajectory information using transition diagrams and fuzzy reasoning [19, 23] ....

M. Tambe. Tracking dynamic team activity. In Proc. Nat. Conf. on Artificial Intelligence, pages 80--87, August 1996.

....the type of rulebased system presented here. Also in the TacAir Soar project (Tambe et al. 1995) the development of intelligent agents for the air combat domain is now coupled with the development of intelligent agents for the soccer domain with special emphasis in the tracking of agents team (Tambe 1996). Also the analysis of the means for coordination of behavior of computer generated forces in TacAir Soar (Laird, Jones Nielsen 1994) offers interesting suggestions for coordination of soccer players. In both the soccer domain and the air combat domain coordination between agents is essential, ....

Tambe, M. 1996 Tracking Dynamic Team Activity, in Proceedings of AAAI96, Portland, Oregon.

....for the algorithm) Heuristics and external knowledge may be used to eliminate paths (hypotheses) which are deemed inappropriate indeed such heuristics will be explored shortly. RESL s basic approachisvery similar to previous work in reactive plan recognition (Rao, 1994) and team tracking (Tambe, 1996), which have been used successfully in the ModSAF domain, and share many of RESL s properties. However, RESL adds belief inference capabilities which are used in the diagnosis process, discussed below (Section 3.4) Figure 4 gives a simpli ed presentation of the plan hierarchies for a variation ....

Tambe, M. (1996). Tracking dynamic team activity. In Proceedings of the National Conference on Articial Intelligence (AAAI).

....called Eavesdropper, which addresses the above difficulties. Eavesdropper uses multiagent plan recognition to monitor agent organizations based on their routine communications. While previous work in multi agent plan recognition has either focused on exploiting explicit teamwork reasoning, e.g. (Tambe 1996), or explicitly reasoning about uncertainty when recognizing multiagent plans, e.g. Intille Bobick 1999) a key novelty in Eavesdropper is that it effectively blends these two threads together. Eavesdropper combines the following novel techniques: First, it uses an efficient, linear time ....

....hypotheses. The individual models approach, however, represents (implicitly) an exponential number of hypotheses. Therefore, tasks that require enumeration of hypotheses will have exponential running times when using individual models. 6 Related Work Like Eavesdropper, previous work by Tambe (Tambe 1996) also focuses on explicitly using team intentions for inferring team plans from observations. However, Eavesdropper uses a more advanced teamwork model (e.g. it can predict failures based on coordination constraints) and also explicitly reasons about uncertainty, allowing it to answer queries ....

Tambe, M. 1996. Tracking dynamic team activity. In Proceedings of the National Conference on Artificial Intelligence (AAAI).

....for the algorithm) Heuristics and external knowledge may be used to eliminate paths (hypotheses) which are deemed inappropriate indeed such heuristics will be explored shortly. RESL s basic approach is very similar to previous work in reactive plan recognition (Rao, 1994) and team tracking (Tambe, 1996), which have been used successfully in the ModSAF domain, and share many of RESL s properties. However, RESL adds belief inference capabilities which are used in the diagnosis process, discussed below (Section 3.4) Figure 4 gives a simpli ed presentation of the plan hierarchies for a variation ....

Tambe, M. (1996). Tracking dynamic team activity. In Proceedings of the National Conference on Articial Intelligence (AAAI).

No context found.

Tambe, M., 1996b. Tracking dynamic team activity.

No context found.

M. Tambe. Tracking dynamic team activity. In Proc. AAAI, 1996.

No context found.

M. Tambe. Tracking dynamic team activity. In Proceedings 13th National Conference on Artificial Intelligence (AAAI-96), pages 80--87, 1996.

No context found.

Milind Tambe, "Tracking Dynamic Team Activity", National Conference on Artificial Intelligence (AAAI96), 1996.

No context found.

Tambe, M.: Tracking dynamic team activity, Proceedings 13th National Conference on Artificial Intelligence (AAAI-96), 1996.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC