CMSC 491M/691M - Spring 2003

Discussion Questions for Class #24, April 23

Reading: Tambe, "Agent Architectures for Flexible, Practical Teamwork" and "Towards Flexible Teamwork."

Reading strategy: Read the shorter paper in detail. In the longer paper, concentrate on Section 1 (Introduction) and Section 2 (Illustrative Domains and Motivation), which give a more detailed analysis of the motivation and issues underlying STEAM. Read Section 3 in detail to be able to discuss the last question below. It isn't necessary to read Sections 4-7 unless you are having trouble understanding the shorter paper and want more details of the algorithms. (Hint: PGP is mentioned in the Related Work section; see third question below.) Read Section 8 to see where STEAM is headed in the future.

• Summarize the three key ideas of STEAM beyond the theoretical framework of joint intentions:
  1. Team synchronization
  2. Monitoring of joint intentions and repair
  3. Decision-theoretic methods for determining what and when to communicate
• The key to team synchronization is establishing joint commitment, which in turn entails that the agents behave in certain ways in response to external events. Give some examples of external events that would trigger action in the context of an established joint commitment.
• STEAM formulates a top-level "team plan," which individual agents expand further to construct their individual plans. The individual plans are then executed in the context of the team plan. In the team planning phase, "all team members must simultaneously select OP to establish a joint intention." This phase requires that there is a team leader who develops the team plan. This is a very different model from the PGP distributed planning system. Discuss the tradeoffs (benefits and disadvantages) of these two models.
• The operator hierarchy in STEAM ("hierarchical reactive plans") represents a somewhat limited notion of planning. A more general planning system might use hierarchical task network planning to construct new strategies on the fly, do more sophisticated situation analysis and variable binding, include contingency plans in case of action failures, and/or use probabilistic models to project likely outcomes and select the best alternatives.
  • How would these planning extensions improve the performance and/or increase the complexity of the overall STEAM system?
  • What kinds of problems might arise in the joint planning activity if individual planners incorporated these extensions?
• What triggers a plan repair in STEAM? What are the repair actions that can be taken?
• What does "decision-theoretic communication selectivity" mean in the context of STEAM? How is the value of a communication action decided?
• What are the influences from SharedPlans and joint intentions theory in STEAM?