Gotta Catch 'Em All! Sequence Flaws in CEGAR for Classical Planning

Martín Pozo, Álvaro Torralba, Carlos Linares López

Abstract: Counterexample-Guided Abstraction Refinement (CEGAR) is a prominent technique to generate Cartesian abstractions for guiding search in cost-optimal planning. The core idea is to iteratively refine the abstraction, by finding a flaw in the current optimal abstract plan. Previous works find only a single flaw, by executing the abstract plan in the concrete state space and stopping when such execution cannot be continued. We show, however, that many flaws can be identified on a single plan. To that end, we introduce sequence flaws, which execute the plan in a Cartesian relaxation of the task to characterize issues beyond the first flaw found along its execution. This greatly increases the flexibility of CEGAR regarding how to refine the abstraction. Our experiments show that a high number of sequence flaws exist in most abstract plans across existing benchmarks. We observe that the selected flaw has a high impact on the resulting heuristic, opening new research opportunities for better selection strategies.