Nathanael Hyafil and Fahiem Bacchus
We present a new algorithm for the conformant probabilistic planning problem. This is a planning problem in which we have probabilistic actions and we want to optimize the probability of achieving the goal, but we have no observations available to us during the course of the plan’s execution. Our algorithm is based on a CSP encoding of the problem, and a new more efficient caching scheme. The result is a gain in performance of several orders of magnitude over previous AI planners that have addressed the same problem. We also compare our algorithm to algorithms for decision theoretic planning. There our algorithm is faster on small problems but does not scale as well. We identify the reasons for this, and show that the two types of algorithms are able to take advantage of distinct types of problem structure. Finding an algorithm that can lever both types of structure simultaneously is posed as an interesting open problem.