Action Model Learning from Noisy Traces: a Probabilistic Approach

This repository contains the official code of the Noisy Offline Learning of Action Models (NOLAM) algorithm.

Installation

pip install nolam

Example usage

from nolam.algorithm.Learner import Learner
noise_rate = 0.1
model = Learner().learn('path/to/domain.pddl', ['path/to/trace0', 'path/to/trace1'], e=noise_rate)
print(model)

Custom domain learning

The NOLAM algorithm can be run for learning from traces with noisy states with an observation noise varying from 0 to 1. For running NOLAM on a custom domain, you need to provide an input domain file 'path/to/domain.pddl', a list of plan trace files ['path/to/trace0', 'path/to/trace1', etc.], and the (possibly estimated) observation noise. The input planning domain must contain the predicates, object types, and operator signatures. Note NOLAM does not yet exploit input knowledge in terms of preconditions and effects, hence providing such input domain knowledge does not currently affect the learning process. NOLAM can learn a planning domain from plan traces of different environments (e.g. it is possible to learn a planning domain from small environments and exploit the learned domain in large environments).

Citations

If you find this repository useful, please consider citing the related paper.

@article{lamanna2024action,
  title={Lifted Action Models Learning from Partial Traces},
  author={Lamanna, Leonardo and Serafini, Luciano},
  booktitle={Proceedings of the International Conference on Automated Planning and Scheduling},
  volume={34},
  pages={342--350},
  year={2024}
}

License

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