Utilizing Structured Representations and CSP's in Conformant Probabilistic Planning

A CSP based algorithm for the conformant probabilistic planning problem (CPP) has been presented by Hyafil & Bacchus. Although their algorithm displayed some interesting potential when compared with traditional POMDP algorithms, it was developed using a “flat” representation. In this work we revisit this algorithm and develop a version that utilizes a structured representation. The structured representation can be exponentially more efficient than the flat representation when dealing with the structured problems that are typical in AI. Our new structured version of their algorithm allows us to demonstrate that the CSP approach to CPP can be much more effective than traditional POMDP algorithms for an interesting range of problems. This is contrary to previously presented results, and makes the application of CSP techniques to decision theoretic planning a promising area for further work.