opt.th • [ ![Build Status] travis-image ] travis [ license-image ] license

Torch library for optimization.

---

You may also be interested in:

• Torch Gurobi bindings at bamos/gurobi.torch.
• Torch ECOS bindings at bamos/ecos.torch.
• bamos/opt.py: Python library for optimization.

---

Algorithms

Spectral Projected Gradient (SPG)

cvx-optim.spg implements the Spectral Projected Gradient (SPG) method [1,2] for convex-constrained optimization.

Projected Gradient Descent (PGD)

cvx-optim.pgd implements projected gradient descent (PGD) for convex-constrained optimization with the proximal algorithm described in Section 4.3 of [3].

Bundle Method (in progress)

Algorithm 1 of [4].

Installation

After installing Torch, cvx-optim.torch can be installed with:

1. luarocks install https://raw.githubusercontent.com/bamos/cvx-optim.torch/master/cvx-optim-scm-1.rockspec
2. Cloning this repo and running luarocks make.

Tests

After installing the library with luarocks, our tests in test.lua can be run with th test.lua.

Licensing

This repository is Apache-licensed.

References

[1] Birgin, Ernesto G., José Mario Martínez, and Marcos Raydan.
"Algorithm 813: SPG—software for convex-constrained optimization."
ACM Transactions on Mathematical Software (TOMS) 27.3 (2001): 340-349.
http://www.ime.usp.br/~egbirgin/publications/bmr2.pdf

[2] Birgin, Ernesto G., José Mario Martínez, and Marcos Raydan.
"Nonmonotone spectral projected gradient methods on convex sets."
SIAM Journal on Optimization 10.4 (2000): 1196-1211.
http://epubs.siam.org/doi/pdf/10.1137/S1052623497330963

[3] Parikh, Neal, and Stephen P. Boyd.
"Proximal Algorithms."
Foundations and Trends in optimization 1.3 (2014): 127-239.
http://www.web.stanford.edu/~boyd/papers/pdf/prox_algs.pdf

[4] Smola, Alex J., S.V.N. Vishwanathan, and Quoc V. Le
"Bundle methods for machine learning."
Advances in Neural Information Processing Systems. 2007.

http://machinelearning.wustl.edu/mlpapers/paper_files/NIPS2007_470.pdf