[Paper Review] Scaling-up Generalized Planning as Heuristic Search with Landmarks

Javier Segovia‐Aguas, Sergio Jiménez+2 more|arXiv (Cornell University)|May 10, 2022

AI-based Problem Solving and Planning4 citations

TL;DR

This paper introduces PGP(f_LM), a novel heuristic search algorithm for generalized planning that combines progressive instance processing with a landmark-based heuristic. By leveraging sub-goal information not explicitly in planning instances, it reduces search plateaus and accelerates convergence, outperforming state-of-the-art methods in both speed and solution quality.

ABSTRACT

Landmarks are one of the most effective search heuristics for classical planning, but largely ignored in generalized planning. Generalized planning (GP) is usually addressed as a combinatorial search in a given space of algorithmic solutions, where candidate solutions are evaluated w.r.t.~the instances they solve. This type of solution evaluation ignores any sub-goal information that is not explicit in the representation of the planning instances, causing plateaus in the space of candidate generalized plans. Furthermore, node expansion in GP is a run-time bottleneck since it requires evaluating every child node over the entire batch of classical planning instances in a GP problem. In this paper we define a landmark counting heuristic for GP (that considers sub-goal information that is not explicitly represented in the planning instances), and a novel heuristic search algorithm for GP (that we call PGP) and that progressively processes subsets of the planning instances of a GP problem. Our two orthogonal contributions are analyzed in an ablation study, showing that both improve the state-of-the-art in GP as heuristic search, and that both benefit from each other when used in combination.

Motivation & Objective

• To address the challenge of search plateaus in generalized planning (GP) caused by heuristic functions that ignore sub-goal information not explicitly represented in planning instances.
• To reduce the runtime bottleneck in GP caused by evaluating every candidate solution across the entire batch of planning instances.
• To adapt classical planning landmark heuristics—originally for state-space search—to the solution-space search in GP.
• To design a novel heuristic search algorithm, PGP, that progressively evaluates candidate plans on subsets of instances, improving scalability and search efficiency.

Proposed method

• Proposes a landmark counting heuristic for GP (f_LM) that extracts sub-goal information—such as necessary actions and state transitions—from the domain representation, even when not explicitly stated in the goal conditions.
• Introduces PGP, a heuristic search algorithm that processes planning instances incrementally rather than evaluating all at once, reducing node expansion cost and avoiding early stagnation.
• Uses landmark graphs derived from classical planning to identify essential actions (e.g., reaching all corridor cells, holding the key) that must be part of any valid solution.
• Combines the landmark heuristic f_LM with progressive evaluation in PGP to guide search in the space of planning programs, improving convergence and reducing search space exploration.
• Employs a best-first search strategy where nodes are prioritized based on f_LM, with dynamic subset processing to balance accuracy and computational cost.
• Adapts classical planning techniques like landmark abstraction and relaxation to the GP setting, enabling transfer of proven heuristics to solution-space search.

Experimental results

Research questions

• RQ1 Can landmark-based heuristics, effective in classical planning, be adapted to improve heuristic search in generalized planning?
• RQ2 Does progressive evaluation of candidate plans on subsets of instances reduce search plateaus and improve scalability in GP?
• RQ3 How do the landmark heuristic and progressive evaluation interact when used together in GP as heuristic search?
• RQ4 To what extent does combining landmark heuristics and progressive processing outperform existing state-of-the-art GP methods?

Key findings

• The landmark counting heuristic f_LM significantly reduces search plateaus by incorporating implicit sub-goal information not present in goal conditions.
• PGP's progressive processing of instances reduces the computational cost of node evaluation, avoiding the full-batch evaluation bottleneck.
• The ablation study shows that both f_LM and PGP individually outperform BFGP, the prior state-of-the-art in GP as heuristic search.
• When combined, f_LM and PGP yield a synergistic improvement, achieving faster convergence and higher solution quality than either component alone.
• The method scales effectively to larger planning problems, demonstrating robustness across diverse generalized planning benchmarks.
• The approach enables transfer of classical planning techniques like landmarks to GP, opening pathways for integrating other classical heuristics such as helpful actions or preferred operators.