We are happy to announce the first release of the LMTR suite for unconstrained optimization. This package contains limited memory trust-region and line-search algorithms implemented in MATLAB. The algorithms are described in "On Efficiently Combining Limited Memory and Trust-Region Techniques", Mathematical Programming Computation (2017) Vol. 9, no 1, pp. 101-134, https://doi.org/10.1007/s12532-016-0109-7.
To download the package, you may want to follow the link http://gratton.perso.enseeiht.fr/LBFGS/index.html
The package contains the following algorithms: - LBFGS_TR.m. Limited memory line-search algorithm L-BFGS that takes a trial step along the quasi-Newton direction inside the trust region; - LBFGS_MTBT.m. Limited memory line-search algorithm L-BFGS based on the Moré-Thuente line search and the initial step is obtained using backtrack; - LBFGS_MT.m. Limited memory line-search algorithm L-BFGS based on the More-Thuente line search; - LMTR_BWX_MS.m. Limited memory trust-region algorithm BWX-MS. It applies the Moré-Sorensen approach for solving the TR subproblem defined in the Euclidean norm. It is a modified version of the algorithm by Burke et al.; - LMTR_EIG_MS_2_2.m. Limited memory trust-region algorithm EIG-MS(2,2) using the eigenvalue-based (2,2)-norm, with the Moré-Sorensen approach for solving a low-dimensional TR subproblem; - LMTR_EIG_inf_2.m. Limited memory trust-region algorithm EIG(inf,2) using the eigenvalue-based (inf,2)-norm, with the exact solution to the TR subproblem in closed form.