Symbolic planning in belief space

Abstract

SASY (Scalable and Adjustable SYmbolic) Planner is a flexible symbolic planner which searches for a satisfying plan to a partially observable Markov decision process, or a POMDP, while benefiting from advantages of classical symbolic planning such as compact belief state expression, domain-independent heuristics, and structural simplicity. Belief space symbolic formalism, an extension of classical symbolic formalism, can be used to transform probabilistic problems into a discretized and deterministic representation such that domain-independent heuristics originally created for classical symbolic planning systems can be applied to them. SASY is optimized to solve POMDPs encoded in belief space symbolic formalism, but can also be used to find a solution to general symbolic planning problems. We compare SASY to two other POMDP solvers, SARSOP and POMDPX_NUS, and define a new benchmark domain called Elevator.