|
CSE Home | AI Home | About CSE | Search | Contact Info |
|
The CBURIDAN PlannerOverviewThe BURIDAN planner introduced actions whose outcomes are probabilistically uncertain. Under the semantics of BURIDAN, it is assumed that the executing agent would not know which outcome occured, and the agent would execute the same sequence of actions regardless of the outcome of previous actions. Clearly, this is not a realistic model of action execution: we generally do get some feedback from actions we perform, and we expect an intelligent agent to use that feedback to tailor its further behavior. Feedback is nearly always incomplete (it does not tell us the entire world state) and often prone to error, thus requiring the agent to reason about the probable true state of the world given the feedback they have received. CBURIDAN extends the BURIDAN action representation to add a classic Bayesian interpretation of information-gathering: actions return "observation labels" whose interpretation is given by the conditional probability that that label would be returned given the possible world-states. Unlike some interpretations of information-gathering, CBURIDAN does not distinguish between "actions" and "observations": any action may both affect the state of the world, and return an observation label. CBURIDAN also extends BURIDAN's plan representation to allow the execution of some actions to be contingent on the outcome of previous actions. For example, a plan might state that a "dry-gripper" action was to be executed only if a previous attempt to pick up a block failed. CBURIDAN plans are "conditional plans", but without the bifurcatinng branching structure of most conditional planners. Publications
|
Computer Science & Engineering University of Washington Box 352350 Seattle, WA 98195-2350 (206) 543-1695 voice, (206) 543-2969 FAX [comments to Dan Weld] |