madupite is a high-performance C++ library with a Python interface designed for solving large-scale Markov Decision P
8000
rocesses (MDPs) using Distributed Inexact Policy Iteration (iPI)1. Leveraging distributed sparse linear solvers from PETSc, madupite efficiently handles the computational complexity associated with large-scale MDPs.
- Scalable MDP Solver: Efficiently solve MDPs with large state and action spaces using distributed computation.
- Python and C++ APIs: Access the power of
madupitethrough both Python and C++, depending on your performance needs. - Distributed Computing: Integrates with PETSc and MPI for distributed computing on multi-core and cluster environments.
After cloning the repository, you can install the package using the following commands.
conda env create -f environment.yml
conda activate madupiteenv
pip install .Note that the conda environment file is limited to Linux for now. After installation, run python examples/install/main.py or mpirun -n <number of ranks> python examples/install/main.py to check if the installation was successful.
We recommend using the provided conda environment. If you prefer to rely on your own environment, make sure to install the required dependencies listed in environment.yml. We refer to the documentation for more information on the installation process.
Solve any stochastic optimal control problem formulated as a MDP using madupite! For example, the fastest way out of a maze or the optimal control input to balance an inverted pendulum:
Examples on how to use madupite can be found in the documentation (tutorials and examples) as well as in the examples directory.
In short: The transition probabilities and stage costs can either be loaded from files or generated using functions in the code (or a combination of both):
- See
examples/tutorial/ex1.pyfor an example on how to generate the required data using function in the code. - See
examples/tutorial/ex2.pyfor an example on how to load the required data from files as well asexamples/tutorial/ex2_data_generation.pyto see how the required files can be generated from NumPy or SciPy arrays.
The tutorials also provides more detail on the data layout and how to preallocate sparse matrices for improved performance.
Contributions to madupite are welcome! Please see the guidelines file CONTRIBUTING.md for more information.
madupite is distributed under the MIT License. See the LICENSE file for more information.
madupite was developed at the Automatic Control Laboratory at ETH Zurich. This work was supported by the European Research Council under the Horizon 2020 Advanced under Grant 787845 (OCAL) and by the SNSF through NCCR Automation (Grant Number 180545).
If you use madupite in your research, please cite it.
@article{Gargiani2025,
doi = {10.21105/joss.07411},
url = {https://doi.org/10.21105/joss.07411},
year = {2025},
publisher = {The Open Journal},
volume = {10},
number = {108},
pages = {7411},
author = {Matilde Gargiani and Philip Pawlowsky and Robin Sieber and Václav Hapla and John Lygeros},
title = {madupite: A High-Performance Distributed Solver for Large-Scale Markov Decision Processes},
journal = {Journal of Open Source Software}
}
Footnotes
-
Gargiani, M.; Sieber. R.; Balta, E.; Liao-McPherson, D.; Lygeros, J. Inexact Policy Iteration Methods for Large-Scale Markov Decision Processes. https://arxiv.org/abs/2404.06136. ↩