Learning to Improve Uncertainty Handling in a Hybrid Planning System

Abstract

Weaver is a hybrid planning algorithm that can create plans in domains that include uncertainty, modelled either as incomplete knowledge of the initial state of the world, of the effects of plan steps or of the possible external events. The plans are guaranteed to exceed some given threshold probability of success. Weaver creates a Bayesian network representation of a plan to evaluate it, in which links corresponding to sequences of events are computed with Markov models. As well as the probability of success, evaluation produces a set of flaws in the candidate plan, which are used by the planner to improve it. We describe a learning method that generates control knowledge compiled from this probabilistic evaluation of plans. The output of the learner is search control knowledge for the planning domain that helps the planner select alternatives that have previously lead to plans with high probability of success. The learned control knowledge is incrementally refined by a combined deductive and inductive mechanism.