Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Safe Learning of PDDL Domains with Conditional Effects -- Extended Version

Argaman Mordoch, Enrico Scala, Roni Stern, Brendan Juba

TL;DR

This work introduces Conditional-SAM, the first algorithm for learning safe PDDL action models in domains with conditional and universal effects. It provides theoretical guarantees on safety, space/time complexity, and sharp sample-complexity bounds for fixed antecedent size, plus extensions to lifted and universally quantified effects. Empirically, Conditional-SAM yields learned models that allow planners to solve test problems in multiple domains, with near-perfect semantic recall of preconditions and exact match of effects to the real model when applicable. The results demonstrate practical applicability while highlighting challenges in domains with large antecedent sizes or complex universal preconditions, guiding future improvements in scalability and expressivity. Overall, the approach offers a principled, safety-preserving path to automatic action-model induction in expressive planning languages.

Abstract

Powerful domain-independent planners have been developed to solve various types of planning problems. These planners often require a model of the acting agent's actions, given in some planning domain description language. Manually designing such an action model is a notoriously challenging task. An alternative is to automatically learn action models from observation. Such an action model is called safe if every plan created with it is consistent with the real, unknown action model. Algorithms for learning such safe action models exist, yet they cannot handle domains with conditional or universal effects, which are common constructs in many planning problems. We prove that learning non-trivial safe action models with conditional effects may require an exponential number of samples. Then, we identify reasonable assumptions under which such learning is tractable and propose SAM Learning of Conditional Effects (Conditional-SAM), the first algorithm capable of doing so. We analyze Conditional-SAM theoretically and evaluate it experimentally. Our results show that the action models learned by Conditional-SAM can be used to solve perfectly most of the test set problems in most of the experimented domains.

Safe Learning of PDDL Domains with Conditional Effects -- Extended Version

TL;DR

This work introduces Conditional-SAM, the first algorithm for learning safe PDDL action models in domains with conditional and universal effects. It provides theoretical guarantees on safety, space/time complexity, and sharp sample-complexity bounds for fixed antecedent size, plus extensions to lifted and universally quantified effects. Empirically, Conditional-SAM yields learned models that allow planners to solve test problems in multiple domains, with near-perfect semantic recall of preconditions and exact match of effects to the real model when applicable. The results demonstrate practical applicability while highlighting challenges in domains with large antecedent sizes or complex universal preconditions, guiding future improvements in scalability and expressivity. Overall, the approach offers a principled, safety-preserving path to automatic action-model induction in expressive planning languages.

Abstract

Powerful domain-independent planners have been developed to solve various types of planning problems. These planners often require a model of the acting agent's actions, given in some planning domain description language. Manually designing such an action model is a notoriously challenging task. An alternative is to automatically learn action models from observation. Such an action model is called safe if every plan created with it is consistent with the real, unknown action model. Algorithms for learning such safe action models exist, yet they cannot handle domains with conditional or universal effects, which are common constructs in many planning problems. We prove that learning non-trivial safe action models with conditional effects may require an exponential number of samples. Then, we identify reasonable assumptions under which such learning is tractable and propose SAM Learning of Conditional Effects (Conditional-SAM), the first algorithm capable of doing so. We analyze Conditional-SAM theoretically and evaluate it experimentally. Our results show that the action models learned by Conditional-SAM can be used to solve perfectly most of the test set problems in most of the experimented domains.
Paper Structure (14 sections, 5 theorems, 6 equations, 3 figures, 3 tables, 4 algorithms)

This paper contains 14 sections, 5 theorems, 6 equations, 3 figures, 3 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Problem Definition and Assumptions
  4. Approach
  5. Conditional-SAM Algorithm
  6. Theoretical Analysis
  7. Learning Lifted Action Models
  8. Learning Effects with Universal Quantifiers
  9. Experimental Results
  10. Evaluation Metrics
  11. Results
  12. Supporting Disjunctive Antecedents
  13. Related Work
  14. Conclusions and Future Work

Key Result

Theorem 3.1

The action model $M'$ learned by Conditional-SAM is safe w.r.t. the action model that generated the input trajectories $\mathcal{T}$.

Figures (3)

  • Figure 1: Parts of the action stop from Miconic domain that contains universally conditional effects.
  • Figure 2: Solving statistics of Satellite, CityCar, Briefcase and Miconic domains.
  • Figure 3: An action representing a flu medicine with a conditional effect.

Theorems & Definitions (16)

  • Definition 1: Safe Action Model
  • Definition 2: Conditional-SAM Inductive Rules
  • Example 1
  • Example 2
  • Theorem 3.1
  • proof
  • Theorem 4.1
  • proof
  • Theorem 4.2
  • proof
  • ...and 6 more