Online Learning of Action Models for PDDL Planning

Refactored implementation of the Online Learning of Action Models (OLAM) algorithm, presented at IJCAI 2021 (paper).

OLAM learns PDDL action models online — interleaving planning and execution — without requiring pre-collected trajectories. Starting from a domain skeleton (predicates and operator signatures with empty preconditions and effects), it incrementally refines the action model as the agent acts in a simulated environment.

The original code used for the paper experiments is preserved at the ijcai-2021 branch.

Installation

From PyPI

pip install olam

From AMLGym

pip install amlgym

Please refer to OLAM integration in AMLGym for a usage example.

For developers

Clone the repository and install in editable mode:

git clone https://github.com/LamannaLeonardo/OLAM.git
cd OLAM
pip install -e .

Quick Start

from unified_planning.io import PDDLReader
from unified_planning.shortcuts import SequentialSimulator
from amlgym.benchmarks import get_domain_path, get_problems_path
from amlgym.util.util import empty_domain

from olam.OLAM import OLAM

# Instantiate a simulated environment from a PDDL domain and problem
domain = 'blocksworld'
domain_ref_path = get_domain_path(domain)
problem_path = get_problems_path(domain, kind='learning')[0]
problem = PDDLReader().parse_problem(domain_ref_path, problem_path)
env = SequentialSimulator(problem=problem)

# Get an input domain path with predicates and operators signature
input_domain_path = empty_domain(domain_ref_path)

# Run the OLAM algorithm
olam = OLAM(input_domain_path)
domain_learned, trajectory = olam.run(env, max_steps=100)

# Print learned domain and produced trajectory
print("##################### Learned domain #####################")
print(domain_learned)

print("################# Generated trajectory ##################")
print(trajectory)

Benchmark Results

OLAM is evaluated on 23 classical planning domains drawn from the IPC benchmark suite. Each domain is paired with a set of 10 problem instances used for incremental online learning. The following table reports for each domain the syntactic precision and recall of the learned model and CPU time (seconds) averaged over all instances.

Citation

If you find OLAM useful for your research, please cite the following paper:

@inproceedings{ijcai2021-566,
  title     = {Online Learning of Action Models for PDDL Planning},
  author    = {Lamanna, Leonardo and Saetti, Alessandro and Serafini, Luciano
               and Gerevini, Alfonso and Traverso, Paolo},
  booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI-21}},
  publisher = {International Joint Conferences on Artificial Intelligence},
  pages     = {4112--4118},
  year      = {2021},
  doi       = {10.24963/ijcai.2021/566},
}

License

This project is licensed under the MIT License — see the LICENSE file for details.

Acknowledgements

Refactored with the help of Ejdis Gjinika.