Reformulation Techniques for Automated Planning: A Systematic Review

Automated planning is a prominent area of Artificial Intelligence, and an important component for intelligent autonomous agents. A cornerstone of domain-independent planning is the separation between planning logic, i.e. the automated reasoning side, and the knowledge model, that encodes a formal representation of domain knowledge needed to reason upon a given problem to synthesise a solution plan. Such a separation enables the use of reformulation techniques, which transform how a model is represented in order to improve the efficiency of plan generation. Over the past decades, significant research effort has been devoted to the design of reformulation techniques. In this paper, we present a systematic review of the large body of work on reformulation techniques for classical planning, aiming to provide a holistic view of the field and to foster future research in the area. As a tangible outcome, we provide a qualitative comparison of the existing classes of techniques, that can help researchers gain an overview of their strengths and weaknesses.

💡 Research Summary

The paper presents a systematic review of reformulation techniques for classical, domain‑independent automated planning. Recognizing that the separation of planning logic from the knowledge model—typically expressed in PDDL—enables the transformation of domain descriptions without altering the planner’s interface, the authors set out to catalogue and compare the myriad methods that have been proposed to improve planner efficiency. A comprehensive literature search covering the period from 1980 to January 2022 was conducted across Google Scholar, Scopus, and the University of Huddersfield’s library, using a two‑tiered keyword strategy (“classical planning reformulation”, “PDDL reformulation”, “domain‑independent reformulation” combined with specific technique terms such as “macro‑operators”, “entanglements”, “action elimination”, “bagged representation”, “action schema splitting”, and “model configuration”). After applying strict inclusion criteria—peer‑reviewed English articles that formally describe at least one reformulation technique, provide empirical evaluation, focus on classical planning, and accept and output PDDL models—the set was narrowed to 54 primary studies.

The authors organize these studies into four principal categories. The first, Macro‑operators/Macro‑actions, aggregates sequences of primitive actions into higher‑level operators, thereby reducing search depth. Notable contributions include Botea & Muñoz‑Ruas (2005) and later works that automate macro discovery via heuristic analysis. The second, Entanglement Resolution/Action Elimination, targets inter‑action dependencies that inflate the state space; techniques such as Helmert’s (2006) entanglement graph prune mutually exclusive actions before search. The third, Bagged Representation/Action Schema Splitting, restructures a single PDDL domain into multiple sub‑domains or groups similar schemas, which can lower memory consumption and improve heuristic evaluation; Fox & Long (2003) exemplify this approach. The fourth, Model Configuration Optimization, revisits the domain definition itself, removing redundant predicates, simplifying type hierarchies, or re‑encoding objects to produce a more compact model.

To illustrate each class, the review employs the well‑known Gripper domain—a robot with two grippers moving several balls between rooms. For macro‑operators, the classic “pick‑move‑drop” triple is collapsed into a single macro action, dramatically shortening the plan length. Entanglement resolution eliminates impossible concurrent grasps, pruning large portions of the search tree. Bagged representation splits the domain into per‑room sub‑problems, allowing planners to solve smaller instances in parallel. Model configuration optimization removes the generic “free” predicate and consolidates room types, yielding a leaner domain file. In each case, the authors report qualitative improvements in search depth, runtime, and memory usage, corroborated by the original studies’ experimental results.

A qualitative comparison matrix evaluates each technique on criteria such as generality, implementation complexity, planner compatibility, pre‑processing overhead, and observed performance gain. Macro‑based methods score high on compatibility and low on preprocessing cost but suffer from limited generality—effective macro design often requires domain‑specific insight. Entanglement and action‑elimination techniques achieve substantial performance boosts in highly interdependent domains but incur significant analysis overhead. Bagged representations excel in large‑scale domains but depend on planners that can handle multiple problem files or hierarchical decomposition. Model configuration offers the deepest structural improvement yet demands extensive domain expertise and may risk altering problem semantics if not carefully validated.

The review also acknowledges gaps in the current literature. Most empirical evaluations are confined to a handful of benchmark planners (e.g., Fast‑Downward, LPG) and standard ICAPS domains, limiting the external validity of reported gains. There is a paucity of standardized benchmarking suites that assess reformulation across diverse planners, hardware platforms, and real‑world applications. Moreover, the integration of machine‑learning techniques—such as reinforcement learning for automatic macro discovery or neural‑guided schema splitting—remains nascent. The authors advocate for the development of open‑source pipelines that automate the end‑to‑end reformulation process, as well as for systematic, cross‑planner studies that quantify trade‑offs between preprocessing time and planning efficiency.

In conclusion, the paper delivers a thorough taxonomy and qualitative synthesis of reformulation techniques for classical planning, highlights their respective strengths and weaknesses, and outlines concrete research directions aimed at automating, standardizing, and extending these methods to broader planning contexts.