Compilation Based Approaches to Probabilistic Planning - Thesis Summary

Models of planning under uncertainty, and in particular, MDPs and POMDPs have received much attention in the AI and DecisionTheoretic planning communities (Boutilier, Dean, and Hanks 1999; Kaelbling, Littman, and Cassandra 1998). These models allow for a richer and more realistic representation of real-world planning problems, but lead to increased complexity. Recently, a new approach for handling certain simple classes of planning under uncertainty was introduced (Palacios and Geffner 2009). This approach works by reducing problems of planning under uncertainty to classical planning problems. The main benefit of this technique is that it allows us to exploit techniques developed in classical planning, and in particular, effective and sophisticated methods for computing heuristic functions. So far, this technique has been shown to be effective for conformant and contingent planning (Albore, Palacios, and Geffner 2009; Shani and Brafman 2011). A related approach was very successful in handling MDPs in the FF-Replan planner (Yoon, Fern, and Givan 2007). The goal of our research is to utilize the approach for probabilistic planning problems. That is, we’d like to be able to represent and reason about probabilities within non probabilistic planning frameworks. To start this research direction, we focus on the problem of conformant probabilistic planning (CPP) with deterministic actions. Although this problem is somewhat narrow, much like conformant planning, it provides a convenient initial step for exploring this research direction. We believe that our techniques can be extended to more general probabilistic planning problems. The best current CPP planner is Probabilistic FF (PFF) (Domshlak and Hoffmann 2007). Probabilistic-FF uses a time-stamped Bayesian Networks (BN) to describe probabilistic belief states. In most benchmarks, PFF’s results were improved by our results.