MOVERS-WORLD presents a challenging planning domain. Each problem requires the actors to move from 3 to 5 boxes of various sizes to the truck. The baseline system initially takes, on average, 38.6 rounds of activity to solve a problem. At any given point in time, an actor's knowledge of the external world is her perceivable environment and a map of the world that she constructs as she goes along. Plans orient the individual actor as to how to proceed (c.f., [70]) but, because of uncertainty in acting, plans are continuously being revised. Actors are assumed to be adaptive planners [4], so if the activity does not unfold as anticipated, the actor's plan will be revised to reflect the ongoing interaction between the actor and her environment. In general, activity may not go as expected for a variety of reasons:
For these reasons, actors will frequently have to adapt, replan, or suspend the current activity. When a plan is completed or abandoned, the actor must select a new set of goals to plan for. At each point in the activity, goals are marked as either `untried', `active', `suspended', `failed', or `completed' and are prioritized during the goal selection process:
A plan is created for a set of goals by either selecting an old plan from memory or creating a plan from scratch. Old plans are stored somewhat abstractly so they require some refinement before they are deployed; our model differs from others in that extracts of execution traces are stored in memory and not plans [18,56,43] or derivational histories [10,73] Others have studied re-using past plans under other conditions [35], including multiple actors [71].
An actor can create a plan from scratch using a given set of STRIPS-like operators [19] in a hierarchical planner (c.f. [63,44]). Depending on their type, different actors have different operations that they can perform. The special purpose operators for the actors in MOVERS-WORLD are:
There are also general purpose operators that any actor can perform, i.e.,
MOVE, WAIT, or SIGN.Each actor maintains a representation of the probability of success for different actions in different contexts. When planning from scratch, the actor uses these probabilities to guide it through the search space [46]. This information represents some of the expectations the individual actor has about the capabilities of herself and her co-participants. The probabilities are incrementally updated, so the actor develops more realistic expectations during the course of action.