Probabilistic Planning with Constraint Satisfaction techniques

Our research’s aim is to explore the use of constraint satisfaction techniques in probabilistic planning. We first focus on two special cases that make different assumptions on the observability of the domain: the conformant probabilistic planning problem (CfPP), where the agent’s environment is not observable, and the contingent probabilistic planning problem (CtPP), where the environment is fully observable. A paper describing some of our work on the first case has been accepted to the technical program of ICAPS 2003 under the title “Conformant Probabilistic Planning via CSPs”. We are currently working on applying similar techniques to CtPP. So far, our research has resulted in exhibiting two independent types of structure that probabilistic planning problems tend to show. Decision theoretic techniques take advantage of state abstraction while our approach, and AI planning techniques in general, rely on reachability properties. Our ultimate goal is to design algorithms that can take advantage of both of these properties, not only in the special cases of CfPP and CtPP but in the general framework of probabilistic planning in partially observable domains.