Learning Features and Abstract Actions for Computing Generalized Plans

Abstract: Generalized planning is concerned with the computation of plans that solve not one but multiple instances of a planning domain. Recently, it has been shown that generalized plans can be expressed as mappings of feature values into actions, and that they can often be computed with fully observable non-deterministic (FOND) planners. The actions in such plans, however, are not the actions in the instances themselves, which are not necessarily common to other instances, but abstract actions that are defined on a set of common features. The formulation assumes that the features and the abstract actions are given. In this work, we address this limitation by showing how to learn them automatically. The resulting account of generalized planning combines learning and planning in a novel way: a learner, based on a Max SAT formulation, yields the features and abstract actions from sampled state transitions, and a FOND planner uses this information, suitably transformed, to produce the general plans. Correctness guarantees are given and experimental results on several domains are reported.

• The paper proposes an automated approach using Max SAT to learn features and abstract actions from state transitions for computing generalized plans.
• The approach transforms learned abstractions into solvable FOND problems, demonstrating effectiveness in experiments but facing scalability limits.
• This method advances generalized planning by automating feature learning, offering a formal contribution and potential links to state representation concepts in machine learning.

Learning Features and Abstract Actions for Computing Generalized Plans

The paper "Learning Features and Abstract Actions for Computing Generalized Plans" addresses the domain of generalized planning, where the aim is to develop strategies that are applicable to multiple instances within a planning domain. The authors present a novel approach to infer features and abstract actions autonomously to compute generalized plans that resolve multi-instance planning problems.

Generalized Planning Formulation

Generalized planning seeks strategies that solve entire classes of problems, rather than individual instances. A notable strategy example in the Blocksworld domain is a plan that iteratively makes a block clear by placing blocks onto the table, irrespective of the initial configuration. Traditionally, generalized plans are mapped directly from observable conditions to actions. However, this paper advances the concept by utilizing a set of boolean and numerical features alongside abstract actions. These abstract actions derive from the transformations of concrete actions based on common feature effects.

Methodology Overview

The primary limitation addressed in this research is the reliance on manually provided features and abstract actions. The authors propose an automated mechanism to learn these components:

• Learning Process: By leveraging a Max SAT (satisfiability) formulation, the paper provides a computational model that learns the features and abstract actions from state transitions in sampled instances.
• Soundness and Completeness: Such abstractions are deemed sound if they ensure that all applicable abstract actions in any given state can be represented by concrete actions. Likewise, completeness requires that all concrete actions can be represented by some abstract counterpart.
• Computational Pipeline: An abstraction is created from validated features and actions, then transformed into a fully observable non-deterministic (FOND) planning problem, solvable via existing FOND planners.

Experimental Results and Implications

The researchers demonstrate the computational model in four distinct planning problems with effective results. They describe instances such as Blocksworld configurations, robot gripper scenarios, and grid-based navigation problems involving reward collection. The experiments indicate practical scalability challenges attributed mainly to the size of SAT encodings and the dimensions of candidate features, necessitating further refinement for broader applications.

Discussion

While promising, this approach faces computational constraints related to the density and diversity of candidate features and SAT theory size, emphasizing the need for enhancements in efficiently handling larger-scale planning domains. Additionally, the feature composition from primitive predicates highlights a potential limitation; a deeper investigation into feature grammar's ability to uniformly capture necessary planning dimensions is warranted.

Potential and Theoretical Contributions

Beyond practical planning, the authors speculate on potential intersections with machine learning concepts like dimensionality reduction and embedding, posing challenges to extend these methods to contexts such as video game state analysis without predefined structure. Future research could focus on learning meaningful state embeddings that maintain abstraction properties essential for generalized planning efficacy.

This paper contributes formal representations and a practical model for generalized planning automation—pivoting from manual feature provisioning to semi-autonomous learning processes—relying on soundness and completeness theory to ensure resulting plans are robust and valid across the generalized problem instances. This formulation presents substantial implications for AI development, offering scalable planning solutions across diverse domains under proper computational management.