Research map / Learning theory and algorithms / Combinatorial and search optimization
Learning Safe Action Models with Partial Observability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PI-SAM and EPI-SAM, two PDDL action-model learners that remain safe (plans are executable and achieve goals) under partially observable trajectories, with sample-complexity guarantees under a bounded concealment assumption.
Problem
Learning PDDL action models from partially observed trajectories (masked fluents) usually yields models that may produce unsafe plans; existing safe learners (SAM) require fully observed states.
Approach
Extend SAM learning rules to apply only when literals are observed in both pre- and post-states (PI-SAM), and extend further with sub-trajectory reasoning compiled into a conformant planning problem with conditional effects (EPI-SAM), which is the strongest safe formulation.
Key results
- PI-SAM achieves perfect empirical precision for preconditions and effects across all IPC domains tested, while FAMA's precision falls below 1 in several domains (e.g., 0.77 for Transport preconditions at η=0.3).
- On Blocks with η=0.3, PI-SAM reaches R(pre)=0.90 and P(pre)=1.00 across 3, 5, and 7 trajectories, while FAMA's precision stays at 0.87–0.90.
- Number of transitions needed to learn preconditions scales inversely with masking-related probability (e.g., Sokoban: 6 at full observability vs 55 at η=0.3 and 563 at η=0.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PI-SAM instead of FAMA | Action model learning under partial observability · Blocks, Depot, Ferry, Floortile, Transport; η = 0.3 | P(pre) | Helped |
| Use PI-SAM instead of FAMA | Action model learning under partial observability · Floortile and Transport; η = 0.3 | R(pre) | Helped |
| Use PI-SAM instead of FAMA | Action model learning under partial observability · Blocks, Floortile, Transport; η = 0.3 | R(eff) | Helped |
Surprising findings
- PI-SAM only applies SAM rules to literals observed in both pre- and post-states, which is counter-intuitive but necessary for safety.
- FAMA's performance may decrease as more input trajectories are given, while PI-SAM's does not.
Limitations
- Assumes deterministic actions, conjunctive preconditions/effects, and noiseless observations.
- Without a bounded concealment assumption, learning a non-trivial safe model under partial observability is impossible.
- Experiments are limited to IPC domains with random masking and a 60-second time limit.
Experimental setup
Datasets: IPC domains (Blocks, Depot, Ferry, Floortile, Gripper, Hanoi, Npuzzle, Parking, Sokoban, Transport)
Benchmarks: IPC domains Blocks, Depot, Ferry, Floortile, Gripper, Hanoi, Npuzzle, Parking, Sokoban, Transport
Metrics: empirical precision (pre, eff) empirical recall (pre, eff) runtime (sec) number of transitions to learn preconditions
Baselines: FAMA EPI-SAM*
Builds on: SAM learning FAMA Michael's learning from partial information
Closest papers by meaning
- Learning Planning Domains from Non-redundant Fully-Observed Traces: Theoretical Foundations and Complexity Analysis (AAAI 2024 · similarity 0.87)
- PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty (NeurIPS 2026 · similarity 0.86)
- Provable Representation with Efficient Planning for Partially Observable Reinforcement Learning (ICML 2024 · similarity 0.85)
- Dealing with Numeric and Metric Time Constraints in PDDL3 via Compilation to Numeric Planning (AAAI 2024 · similarity 0.84)
- Learning Planning Abstractions from Language (ICLR 2024 · similarity 0.84)
- Leveraging Environment Interaction for Automated PDDL Translation and Planning with Large Language Models (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.