Les conflits entre agents

Orlando, Florida, July 18, 1999

• Description
• Workshop summary
• Workshop committee
• More on conflicts
• Further information

Agents' conflicts arise for different reasons, involve different concepts, and are dealt with in different ways, depending on the kind of agents and on the domain where they are considered. For example, agents may have conflicting beliefs, conflicting goals, or may have to share limited resources. Conflicts can be expressed as mere differences, or as contradictions, or even as social conflicts (e.g. aggression, fighting). They may be avoided, solved, kept, or even created deliberately.

Since more and more concern is attached to agents' teamwork and agents' dialogue, conflicts naturally arise as a key issue to be dealt with, not only with application dedicated techniques, but also with more formal and generic tools.

The aim of the workshop is therefore to focus on definitions of agents' conflicts and on their roles within a multiagent system, i.e. how this system may evolve thanks to, despite, or because of conflicts.

AN APPOINTMENT SCHEDULING COMMON SCENARIO
In order to have a common application wherein the workshop participants may present their techniques, we suggested appointment scheduling via personal assistant agents. Some of the participants have used the scenario to work out vivid examples to demonstrate their techniques.

WORKSHOP SUMMARY

Towards conflict resolution in agent teams via argumentation (invited) - M.Tambe and H.Jung
Conflicts: unanticipated resource conflicts, unanticipated action conflicts due to inconsistent beliefs (local information, different interpretations).
Detection: e.g. constraints attached to agents' roles are violated.
Key idea: conflict resolution is a team goal.

Identifying conflicting probabilistic knowledge in multiple knowledge base systems (regular) - D. O'Leary
Conflicts: experts have different views of the world (error in one's judgement, unspecified model).
Detection: difference between probabilities exceeds a given threshold.
Key idea: search for an agreement for well-defined problems (e.g. do I have cancer or not?), but not via averaging.

Matching conflicts: functional validation of agents (regular) - G.Cybenko and G.Jiang
Conflicts: matching conflicts between client and server, service is not adapted.
Detection: mathematical criterion.
Key idea:

Defeasible reasoning between conflicting agents based on VALPSN (regular) - K.Nakamatsu, J.M.Abe and A.Suzuki
Conflicts: contradiction [P and (not P)].
Detection: logical inconsistency.
Key idea: ordinal weighting on formulas.

Analysis of mental attributes for the conflict resolution in multiagent systems (regular) - B.Galitsky
Conflicts:
Detection:
Key idea:

Spatial conflicts among robot agents (invited) -J.Penders
Conflicts: spatial conflicts among robots
Detection: when two robots can observe each other.
Key idea: conflicts among a team of robots generate a derived behaviour (e.g. they follow a given target).

Conflicts in a simple autonomy-based multi-agent system (regular) - F.Chantemargue, P.Lerena and M.Courant
Conflicts:spatial conflicts, concurrent access to physical space, which is a type of resource (e.g. two robots that want to come to the same location).
Detection: a robot that can't come to a given location because another robot is already there will somehow see some kind of conflict: this is reflected by the fact that the robot is perturbed; its trajectory undergoes a modification due to the other robot; the robot has to choose another location, which of course is a second choice location (the primary choice is not physically possible).
Key idea: nothing is explicitely undertaken to eliminate conflicts.

Conflict and interaction uncertainty: two applications in the field (short) - W.Lawless, T.Castelao and C.Abubucker
Conflicts: two different and incommensurable views of the world.
Detection: agents do not agree with each other.
Key idea: make decision based on conflict.

Resolving conflicts in collaborative human-computer interaction (invited) - R. St.Amant
Conflicts: human-human and agent-human conflicts.
Key ideas: human-human conflict handling: explicit problem statements, iterative preference refinement, visibly low-effort information access, procedural guidance; agent-human (HCI) conflict management: user given a restricted set of actions, perceived cost of action reduced if less likely to lead to a conflict, present justification for action or belief, present relevant information about context, provide navigational mechanisms for evaluation and redirection [from the author's slides].

DISCUSSION - Conclusions

• It is easier not to be in conflict than to be in conflict. The former may mean that the agents are not even interacting. The latter supposes that the agents are within the same context.
• A conflict does not seem to be necessarily symmetric: conflict (a, b) does not imply conflict (b, a).
  This is particularly the case for spatial conflicts: when two robots roam in a 2D space for instance, the notion of spatial conflict appears only at the time when a robot attempts to move to the location of the other robot. But the latter does not see this conflict.
• Are conflicts useful? The answer depends on the problem. To be useful, a conflict must be observed.
• What we learn from a conflict depends on the situation. Learning is possible if agents are aware of the conflict.
• Conflicts are positive in certain cases, e.g. they may create specific behaviours (cf Penders's paper), create competition, or stimulate inference.

WORKSHOP COMMITTEE

Cristiano Castelfranchi, CNR, Italy
Laurent Chaudron, Onera-Cert, Toulouse, France (co-chair).
Mark Klein, CCS-MIT, USA
Juergen Mueller, Deutsche Telekom, Germany
Joël Quinqueton, Inria, France
Munindar P. Singh, North Carolina State University, USA
Milind Tambe, University of Southern California, USA
Catherine Tessier, Onera-Cert, Toulouse, France (co-chair).

MORE ON CONFLICTS

Visit the Computational models of conflict management web site!

See what happened at the IJCAI97 workshop on Collaboration, Cooperation and Conflict in Dialogue Systems!

See also what happened at the ECAI'98 workshop Conflicts among agents: avoid or use them?