amlgym 1.0.0

Action model learning benchmarking.

AMLGym: benchmarking action model learning

Official code for benchmark generation and evaluation of action model learning approaches.

Installation

pip install amlgym

Example usage

from amlgym.algorithms import get_algorithm
agent = get_algorithm('OffLAM')
model = agent.learn('path/to/domain.pddl', ['path/to/trace0', 'path/to/trace1'])
print(model)

State-of-the-art Algorithms

AMLGym provides seamless integration with state-of-the-art algorithms for offline learning classical planning domains from an input set of trajectories in the following settings:

1. full observability: SAM [1].
2. partial observability: OffLAM [2].
3. full and noisy observability: NOLAM [3], ROSAME [4].

It is possible to run the above algorithms as of the main.py script, which by default runs SAM, OffLAM, NOLAM, and ROSAME on every domain and associated set of trajectories in benchmarks/trajectories/learning.

[1] "Safe Learning of Lifted Action Models", B. Juba and H. S. Le, and R. Stern, Proceedings of the 18th International Conference on Principles of Knowledge Representation and Reasoning, 2021.

[2] "Lifted Action Models Learning from Partial Traces", L. Lamanna, L. Serafini, A. Saetti, A. Gerevini, and P. Traverso, Artificial Intelligence Journal, 2025.

[3] "Action Model Learning from Noisy Traces: a Probabilistic Approach", L. Lamanna and L. Serafini, Proceedings of the Thirty-Fourth International Conference on Automated Planning and Scheduling, 2024.

[4] "Neuro-symbolic learning of lifted action models from visual traces", X. Kai, S. Gould, and S. Thiébaux, Proceedings of the Thirty-Fourth International Conference on Automated Planning and Scheduling, 2024.

Adding an algorithm

PRs with new or existing state-of-the-art algorithms are welcome:

1. Add the algorithm PyPI package in requirements.txt
2. Create a Python class in algorithms which inherits from AlgorithmAdapter.py and implements the learn method

Evaluation

AMLGym can evaluate a PDDL model by means of several metrics:

1. Syntactic precision and recall
2. Predicted applicability and effects
3. Solvability metrics: False plans and problem solving ratios.

Benchmarking

See the benchmark package for details.

License

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