Non-Deterministic Conformant Planning Using a Counterexample-Guided Incremental Compilation to Classical Planning

Abstract

We address the problem of non-deterministic conformant planning, i.e., finding a plan in a non-deterministic context where the environment is not observable. Our approach uses an unsound but complete reduction from non-deterministic conformant planning to classical planning to find a candidate plan; the validity of this plan is then verified by a SAT solver; if the plan is invalid, the reduction is revised to guarantee that the invalid plan will not be valid in the classical planning problem. This procedure is executed until a valid plan is found, or it is shown that there is no plan. Experiments show that this approach provides a nice trade-off between fast but unsound, and complete but slow approaches.