PySLSQP

The SLSQP algorithm is designed to solve nonlinear programming (NLP) problems. PySLSQP is a Python package that wraps the original SLSQP algorithm implemented in Fortran by Dieter Kraft. While the Fortran code is sourced from scipy.optimize.minimize, PySLSQP extends its functionality by offering new features for further analysis of optimization problems, thereby significantly improving the utility of the original algorithm. The prebuilt packages for various system architectures are available on PyPI for download, thus avoiding the need for the user to compile the Fortran sources.

Some of the additional features provided by PySLSQP include:

• Data Saving: PySLSQP allows you to save optimization data during the optimization process. This can be useful for analyzing the progress of the optimization, for post-processing purposes, or for warm/hot restart of the optimization.

• Warm/Hot Starting: PySLSQP supports warm/hot starting, enabling users to initialize the optimization algorithm with a previously saved solution. This can be particularly useful when you want to restart an optimization from a previous solution or continue an optimization that was terminated for various reasons.

• Live Visualization: PySLSQP provides the capability to visualize the optimization progress in real-time, enabling users to monitor the convergence via the optimality and feasibility measures. This also helps the users understand how the objective function, constraints, and design variables are changing during the optimization.

• Scaling: PySLSQP allows you to independently scale the objective function, constraints, and design variables, separate from their original definitions. Scaling can help improve the convergence behavior of the optimization algorithm and make it more robust.

• More Internal Optimization Variables: PySLSQP provides access to additional internal optimization variables such as optimality, feasibility, Lagrange multipliers, etc. which can be useful for advanced analysis of the optimization problem. In scipy.optimize.minimize, the original callback function returns only the optimization variables, and only for the major iterations.

These additional features make PySLSQP a powerful tool for solving constrained optimization problems in Python. In addition, PySLSQP also supports the estimation of gradients using first-order finite differencing, as in the Scipy version.

Installation

To install the latest release of PySLSQP on PyPI, run on the terminal or command line

pip install pyslsqp

Warning

Precompiled wheels for common Ubuntu, macOS, and Windows architectures are available on PyPI. However, if a wheel for your system's architecture is not available, the above installation will compile the source distribution directly on your machine. In such scenarios, if your Fortran compilers aren't compatible, you may encounter compilation errors. Additional troubleshooting may be required to resolve these errors depending on their specifics.

To install the latest commit from the main branch, run

pip install git+https://github.com/anugrahjo/PySLSQP.git@main

Note that this installation method will compile the Fortran sources locally on your machine. Therefore, we only recommend this method if you are a developer looking to modify the package for your own use case.

To upgrade PySLSQP from an older version to the latest released version on PyPI, run

pip install --upgrade pyslsqp

To uninstall PySLSQP, run

pip uninstall pyslsqp

Testing

To test if the package works correctly and as intended, install pytest using

pip install pytest

and run the following line on the terminal from the project's root directory:

pytest -m "not visualize"

Usage

Most features of the PySLSQP package can be accessed through the optimize function. However, there are some additional utility functions that are available for post-processing. Here is a small optimization example that minimizes x^2 + y^2.

import numpy as np
from pyslsqp import optimize

# `v` represents the vector of optimization variables
def objective(v):
    # the objective function
    return v[0]**2 + v[1]**2

x0 = np.array([1., 1.])
# optimize() returns a dictionary that contains the results from optimization
results = optimize(x0, obj=objective)
print(results)

Note that we did not provide the gradient for the objective function above. In the absence of user-provided gradients, optimize estimates the gradients using first-order finite differencing. However, it is always more efficient for the user to provide the exact gradients. Note also that we did not have any constraints or variable bounds in this problem. Examples with user-defined gradients, constraints, and bounds can be found in the Basic User Guide.

Documentation

For API reference and more details on installation and usage, visit the documentation.

Citation

If you use PySLSQP in your work, please use the following reference for citation:

@article{joshy2024pyslsqp,
  title={PySLSQP: A transparent Python package for the SLSQP optimization algorithm modernized with utilities for visualization and post-processing},
  author={Joshy, Anugrah Jo and Hwang, John T},
  journal={arXiv preprint},
  year={2024},
  doi= {10.48550/arXiv.2408.13420},
}

Bugs, feature requests, questions

Please use the GitHub issue tracker for reporting bugs, requesting new features, or any other questions.

Contributing

We always welcome contributions to PySLSQP. Please refer the CONTRIBUTING.md file for guidelines on how to contribute.

License

This project is licensed under the terms of the BSD 3-Clause "New" or "Revised" License.