Next: 3.5 Coordinating Joint Activity Up: 3. MOVERS-WORLD Previous: 3.3.1 Operator Probability Trees

  
3.4 Communication

MOVERS-WORLD, and the techniques used to learn coordinated procedures by MOVERS-WORLD actors, is shaped by the fact that communication is the central mechanism for cooperation and coordination. Communication is considered `expensive', so it occurs at run-time and is never part of the planning process; this is unlike standard planning systems [15,74]. Actors limit themselves to single requests; transmitting single requests lowers both communication costs and plan merging costs, which can be considerable for distributed, independent actors.

Two actors are said to `cooperate' if they act or work together to achieve some common purpose and they are `coordinated' when their individual actions are appropriately ordered to support cooperative activity. In our framework, communication is the only mechanism whereby actors can determine if they are cooperating. In other words, there are no global structures, such as blackboards (e.g., [50]), for actors to use to determine if they happen to be working on the same goal. Each MOVERS-WORLD problem-solving episode includes some goal(s) that can only be solved by cooperating actors, so communication is an essential part of the community activity.

Communication is used to attempt to establish cooperation when the set of actors working on a goal at a given time is inadequate. Cooperation is not guaranteed during communication since actors have their own decision-making strategies; even if an actor is willing to cooperate, she may be unable to do so. An actor who is unwilling or unable to assist can propose an alternative that the original requester may now contemplate adopting. All actors presently use the same strategy for deciding whether to cooperate: preference is given to the plan that achieves the highest expected number of top-level goals. Despite their common strategy, actors will make different decisions because of differences in their experiences.

Communication between actors in our system is similar to a telephone conversation. Actors do not have to be in the same location to engage in communication. One actor can call another and either establishes a connection, gets a busy signal (only two actors can participate in a single negotiation -- as opposed to contract nets: [16] -- or the other actor refuses to respond. Once a connection is established, communication is handled via request and response frames. After agreeing to cooperate, an actor can opt-out at any time, without obligation to notify other actors (in contrast to [51,32]).

There are three initial request types with fifteen response types, ten of which require a response from the original caller. In some situations, the caller may follow up with a reiteration of the original request, so a single conversation can contain between one and three requests. In practice, not all of the 315 possible sequences of request and response types are possible; in one test run involving over 24,000 conversations, 57 different sequences occurred, of which 14 contained a single request, 30 contained two requests and 13 contained three requests.


Next: 3.5 Coordinating Joint Activity Up: 3. MOVERS-WORLD Previous: 3.3.1 Operator Probability Trees
Last Update: March 10, 1999 by Andy Garland