NLPy / News: Recent posts

There is now an interface to the Harwell 2007 package MA57 for the solution of symmetric indefinite matrices such as those arising in constrained optimization, least squares and symmetric saddle-point problems (such as the Stokes problem). MA57 is a major improvement over MA27 and is free to academics and researchers. I am working on a common interface for both MA27 and MA57.

A pure Python implementation of MINRES is now in NLPy. MINRES solves symmetric systems or linear least-squares problems. The code makes provision for preconditioning and allows the coefficient matrix to be given as an operator.

A Python implementation of LSQR was just added to NLPy. LSQR is Michael Saunders' implementation of a conjugate-gradient-type iterative for sparse linear equations linear least-squares problems. See Michael's page for more information http://www.stanford.edu/group/SOL/software/lsqr.html

NLPy is now entirely based on the new NumPy. Numeric is now considered deprecated and is no longer supported. Please use NumPy v1.0.0 and above.

A new module was added, implementing the preconditioned projected gradient algorithm used to solve EQPs (equality-constrained quadratic programs). A demo file is included

At the same time, LibAmpl, the AMPL Library bindings, became an optional package. If a user wants to use ppcg, they will not necessarily need LibAmpl.

A new optional package, csrch from the Minpack-2 package implementing the More and Thuente linesearch, was added. This linesearch procedure attempts to satisfy the strong Wolfe conditions. Jorge Nocedal's modified version, mcsrch, found in his limited-memory BFGS code was also added, and nlpy now sports a Python LBFGS solver for unconstrained minimization. Only the linesearch is in Fortran. The LBFGS class may also be used to build preconditioners.

NLPy now supports optional packages. The basic requirements have been reduced to a minimum. The packages MA27, ICFS, TRON and GALAHAD are now optional. NLPy may be built without these packages installed. Any optional package may be installed later with minimal rebuild. Check http://nlpy.sf.net/news.html