Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning

Zenodo (CERN European Organization for Nuclear Research)

Abstract

Datasets, trained model checkpoints and planner outputs for the NeurIPS 2026 paper "Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning" (Aichmüller, Ståhlberg, Funkquist, Geffner). The record contains everything behind the paper's coverage tables, plus a snapshot of the code to train and run the policies. Contents:- README.md: layout, how to run a checkpoint, the mapping from folders to paper columns, and how the reported checkpoints were chosen.- manifest.tsv: one row per (paper column, domain, checkpoint), with validation and test results and the paths of the checkpoint and its evaluation.- datasets.zip: the 21 planning domains of the paper (10 from the IPC 2023 learning track, 11 further domains) with their train/validation/test splits in PDDL.- checkpoints.zip: 1,169 trained R-GNN policies. For every paper column and domain, this is the checkpoint whose test coverage is reported, plus the next best checkpoints of its planning round.- evaluations-<column>.zip (one per paper column): the planner output of every checkpoint on the test split, i.e. plans, per-problem summaries and search traces.- code.zip: snapshot of https://github.com/maichmueller/aggregated-delta (training and planning code, prebuilt wheels for CPython 3.12 on Linux x86-64, Dockerfile and Apptainer definition), plus source snapshots of the exact dependency versions. Usage: unzip all archives into one folder, build the container from code/ and plan from that folder (see README.md). Running checkpoints only needs code.zip, datasets.zip and checkpoints.zip. Counting "Solved: yes" in the per-problem summaries of the reported checkpoints reproduces every coverage number of the paper. Licenses: data, checkpoints and planner outputs are released under CC BY 4.0. The code in code.zip is released under GPL-3.0 (see code/LICENSE).

Authors 4

1. Michael Aichmüller Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

2. Simon Ståhlberg Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

3. Martin Funkquist

   Linköping University

   Affiliation as printed

   Linköping University

4. Héctor Geffner Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University