Combining Knowledge Compilation and Search for Conformant Probabilistic Planning

We present a new algorithm for conformant probabilistic planning, which for a given horizon produces a plan that maximizes the probability of success under quantified uncertainty about the initial state and action effects, and absence of sensory information. Recent work has studied systematic search in the space of all candidate plans as a feasible approach to conformant probabilistic planning, but the algorithms proposed require caching of intermediate computations in such a way that memory is often exhausted quickly except for small planning horizons. On the other hand, planning problems in typical formulations generally have treewidths that do not grow with the horizon, as connections between variables are local to the neighborhood of each time step. These existing planners, however, are unable to directly benefit from the bounded treewidth owing to a constraint on the variable ordering which is necessary for correct computation of the optimal plan. We show that lifting such constraint allows one to obtain a compact compilation of the planning problem, from which an upper bound can be efficiently computed on the value of any partial plan generated during search. Coupled with several optimizations, this results in a depth-first branchand-bound algorithm which on the tested domains runs an order of magnitude faster than its predecessors, and at the same time is able to solve problems for significantly larger horizons thanks to its minimal memory requirements.