Code for Block BFGS Methods

Last updated: Jan. 04 , 2017.

Here you can find the code used for the numerical experiments in Section 7 of Block BFGS Methods.
Code: blockbfgs_dist_2.zip
The code was changed for Version 2 of our paper. In addition to changing the setup of the experiments, our quasi-Newton implementation is substantially improved.

The test functions are distributed separately, as some of the files are large. Please consult 'README.txt' in the code package before downloading the tests. You should copy these files over the placeholder folders in the main code package. At the moment, these packages contain the original Matlab environments. In the future, I hope to release the data in an open format.

logreg_tests.zip
Logistic regression problems. These test functions contain data from the datasets distributed through LIBSVM. It also uses several functions from the StochBFGS package of Robert Gower.
logb_tests.zip
Log barrier QPs and LPs using data from István Maros' and Csaba Mészáros' QP collection and William Hager's COAP test problems.
Warning: this file is very large (~630MB) and contains more data than is needed for the tests. You might be better off generating some QPs and LPs of your own to test.
tanh_tests.zip
Hyperbolic tangent loss problems. These test functions contain data from the datasets distributed through LIBSVM.
benchmark_tests.zip
Problems from An unconstrained optimization test functions collection by Neculai Andrei.

You will also need:

minFunc, Mark Schmidt's unconstrained optimization library for Matlab.

The old code from Version 1 can be found here: blockbfgs_dist.zip

Please contact me with any questions or bug reports.
Back to homepage.