Discussion

  We will now discuss the constraints of the task environment that CM must accommodate.

Agents have limited knowledge about each other.

Agents have no a priori knowledge of other agents' capabilities; instead, the other agents are considered black boxes. Since the only way to characterize black boxes is to construct a mapping between inputs and outputs, an agent can only assess abilities based upon the success or failure (the output) of attempted operators (the input).

Communication is necessary.

Communication is essential in a domain such as MOVERS-WORLD for two reasons. First, the impoverished model of agent capabilities is inadequate to use communication-free techniques, such as the analytic approach of [Genesereth, Ginsberg, & Rosenschien1986] and the observation-based approach of [Huber & Durfee1995]. Second, there is neither an alternative to communication to allocate more agents to a goal, nor is there an incentive for agents to take on extra goals. So, when the set of agents working on a goal is inadequate, communication must be used to enlist help. Thus, an approach that treats other agents as just another object in the environment, such as [Sen, Sekeran, & Hale1994], is not appropriate.

Operators and plans will sometimes fail.

Agent operators may not work as expected for a variety of reasons including an incomplete domain model (c.f., the qualification problem) and an incorrect assessment of agent abilities. In addition to operator failure, the reasons that a plan might be abandoned include inaccurate or outdated situation knowledge.

Communicate while executing, not planning.

If communication must occur, it is desirable to limit the burden it places on the agents. If communication is fast relative to cogitation, then we should freely incorporate it into the planner since the benefits would outweigh a marginal slowdown in planning. However, we feel that this is not a reasonable assumption for many domains since processor technology historically increases at a faster rate than communication technology. We therefore limit our communication to very simple requests and responses. We further assume that this limited communication and action take roughly the same amount of time, which is appropriate for the types of actions we consider in our domain.

Since agents might have to replan frequently due to operator or plan failure, and communication is expensive, and any communication establishing cooperation for steps of a plan that are never executed is wasted, communication should occur while executing rather than while planning. Standard DAI planners, even a functionally accurate planner such as [Corkill1979], that use communication during planning are therefore not appropriate. It is possible that if a single agent fails, these planners would involve the whole community in replanning and recommunicating. Likewise, a contract net approach [Davis & Smith1983] would potentially have to constantly revise the bid and task announcements as new information becomes available.