Counterexample-Guided Abstraction Refinement for Pattern Selection in Optimal Classical Planning

Abstract

We describe a new algorithm for generating pattern collections for pattern database heuristics in optimal classical planning. The algorithm uses the counterexample-guided abstraction refinement (CEGAR) principle to guide the pattern selection process. Our experimental evaluation shows that a single run of the CEGAR algorithm can compute informative pattern collections in a fairly short time. Using multiple CEGAR algorithm runs, we can compute much larger pattern collections, still in shorter time than existing approaches, which leads to a planner that outperforms the state-of-the-art pattern selection methods by a significant margin.

Pattern Selection Strategies for Pattern Databases in Probabilistic Planning

“…The hillclimbing configurations tend to perform even better than CEGAR in this setting. However, we point out that in classical planning, Rovner, Sievers, and Helmert (2019) achieved significantly better results when combining the PDBs generated by the disjoint CEGAR algorithm with saturated cost partitioning (Seipp and Helmert 2014;Seipp, Keller, and Helmert 2020), as the patterns produced by CE-GAR are typically too large to be combined using ordinary additivity constraints. As of yet, it is unclear whether or how we can pursue a similar strategy to combine the patterns produced by CEGAR, so it is not yet safe to say that hillclimbing is all in all the better strategy to use for SSP problems.…”

“…In its core, CEGAR iteratively refines an abstraction of the state space until the abstraction satisfies a property of interest or a resource limit is reached. Especially interesting for us is its use in a generation algorithm for pattern collections in classical planning by Rovner, Sievers, and Helmert (2019), using projection as the underlying abstraction type.…”

“…We instantiate this policy-based CEGAR framework by adapting the pattern generation algorithm by Rovner, Sievers, and Helmert (2019), which operates on multiple disjoint projections. In a nutshell, this algorithm starts with the disjoint pattern collection containing only a singleton pattern for each goal variable.…”

“…Here we fill that gap, for both the SSP and the MaxProb setting. We adapt two popular strategies known from classi-cal planning: Counterexample guided abstraction refinement (CEGAR) (Rovner, Sievers, and Helmert 2019) and hillclimbing search in the space of pattern collections (Haslum et al 2007), which we only pursue for SSP problems. We analyze how these methods need to be extended for the probabilistic setting.…”

Pattern Databases for Goal-Probability Maximization in Probabilistic Planning

“…This yields an abstract state space where states that agree on P are not distinguished. PDB heuristics (Haslum et al 2007;Franco et al 2017;Rovner, Sievers, and Helmert 2019) and other more general abstraction heuristics (Helmert et al 2014;Seipp and Helmert 2018) are of paramount importance for effective cost-optimal classical planning.…”