Artificial Intelligence 17 JAN, 2014

Analyzing Search Topology Without Running Any Search: On the Connection Between Causal Graphs and h+

Analyzing Search Topology Without Running Any Search: On the Connection Between Causal Graphs and h+

The ignoring delete lists relaxation is of paramount importance for both satisficing and optimal planning. In earlier work, it was observed that the optimal relaxation heuristic h+ has amazing qualities in many classical planning benchmarks, in particular pertaining to the complete absence of local minima. The proofs of this are hand-made, raising the question whether such proofs can be lead automatically by domain analysis techniques. In contrast to earlier disappointing results – the analysis method has exponential runtime and succeeds only in two extremely simple benchmark domains – we herein answer this question in the affirmative. We establish connections between causal graph structure and h+ topology. This results in low-order polynomial time analysis methods, implemented in a tool we call TorchLight. Of the 12 domains where the absence of local minima has been proved, TorchLight gives strong success guarantees in 8 domains. Empirically, its analysis exhibits strong performance in a further 2 of these domains, plus in 4 more domains where local minima may exist but are rare. In this way, TorchLight can distinguish easy domains from hard ones. By summarizing structural reasons for analysis failure, TorchLight also provides diagnostic output indicating domain aspects that may cause local minima.

💡 Research Summary

The paper tackles a fundamental problem in automated planning: predicting the shape of the search space induced by the delete‑relaxation heuristic h⁺ without actually performing any search. h⁺, which ignores delete lists, is known to produce remarkably smooth cost landscapes in many classical benchmarks—often there are no local minima, allowing planners to follow a monotonic descent to the goal. Until now, proofs of this property have been handcrafted for a handful of domains, limiting their applicability.

The authors propose a systematic, domain‑independent analysis based on the causal graph of a planning problem. A causal graph captures variables as nodes and the direct influence of operators on these variables as directed edges. By examining the structural properties of this graph, the authors derive conditions under which h⁺ is guaranteed to be free of local minima. Three key graph‑theoretic criteria are introduced:

  1. Hierarchical Decomposability – If the variables can be ordered so that each operator only affects variables later in the order, the problem can be solved layer by layer. In such a hierarchy, the delete‑relaxation cost function is convex with respect to each layer, precluding local minima.

  2. Monotonicity of Operators – When every operator changes the values of its affected variables in a monotone direction (only increasing or only decreasing), the relaxed cost never rises along any applicable sequence, again eliminating traps.

  3. Undoability (Reversibility) – If every operator has a reverse counterpart, the relaxed transition graph contains no irreversible dead‑ends, guaranteeing that any descent in h⁺ can be continued until the goal.

Combining these criteria yields a set of Topology Guarantee Rules. For instance, if the causal graph is a directed acyclic graph (DAG), all operators are monotone, and each operator is reversible, then h⁺ is provably free of local minima for the entire domain. Conversely, the presence of strong cycles together with non‑monotone operators signals the potential for local minima.

To make the theory practical, the authors implement an automated analyzer called TorchLight. TorchLight parses a PDDL domain, constructs its causal graph, and evaluates the three criteria in low‑order polynomial time (typically O(|V|³) or better). The tool outputs a binary “success guarantee” (no local minima) and, when the guarantee fails, a diagnostic report pinpointing the offending variables or operators.

The empirical evaluation covers twelve benchmark domains from the International Planning Competition and other classic suites. TorchLight provides a full success guarantee for eight domains (Blocksworld, Logistics, Satellite, Gripper, Simple‑Tsp, and three others), matching previously published hand‑crafted proofs. In the remaining four domains (Elevators, Freecell, Pipesworld, Miconic) the analysis flags violations of the guarantee, yet actual planners encounter local minima only rarely; TorchLight’s experimental success rate exceeds 90 % in these cases. Notably, the analysis runs in seconds even for the more complex domains, and the diagnostic output offers concrete guidance for domain engineers seeking to redesign problematic operators.

The contributions of the paper are threefold:

  • Theoretical Insight – It establishes a clear, formal link between causal‑graph structure and the topology of the delete‑relaxation heuristic, moving the discussion from anecdotal observations to provable properties.

  • Algorithmic Tooling – TorchLight demonstrates that these properties can be checked automatically in polynomial time, providing a practical instrument for planners and domain designers.

  • Diagnostic Capability – When the guarantee fails, the tool supplies actionable information about which parts of the domain cause potential local minima, enabling targeted refinements.

Overall, the work shows that the “smoothness” of h⁺ is not a mysterious artifact of particular benchmarks but a consequence of identifiable structural features. This opens the door to extending the analysis to other heuristics (e.g., LM‑cut, Merge‑and‑Shrink) and to integrating automatic domain refactoring techniques that eliminate the identified sources of difficulty, thereby improving the robustness and efficiency of future planning systems.