- Now returns a set of axes instead of showing the plot, - make sure to create a fresh set of axes at each call.
If a solver consistently dominates the others, its ratio array is constant. This causes no horizontal line to appear on the plot. Fix this by moving the last abscissa to xmax.
LSR1 approximation has been created, but is yet to be tested.
Fixed a small counting error in LBFGS class within matvec calculation.
Fixed code based on a mistake in the derivation of the Hessian of the augmented Lagrangian.
Test of augmented Lagrangian form of NLP passed successfully. Commit includes several bug fixes to mfnlp.py and auglag.py.
Migrated new files from a personal hg repository and updated links to other NLPy files. Also, added a restart function to InverseLBFGS to allow possible restarting of the approximation e.g. in subproblem solves.
Also, fixed naming convention: LBFGS class is simply "LBFGS" instead of "DirectLBFGS"
LBFGS now runs correctly when compared to Inverse LBFGS. Applying matvecs to an arbitrary vector first with inverse, then with direct BFGS approximations yields the original vector.
Created a class in lbfgs.py that computes matrix-vector products with the LBFGS approximate Hessian directly, not its inverse. Also added a dummy class in bqp.py for compatibility with older python versions. (This should be moved to the utilities section of the code for broader applicability.)
Make parameter updates and infeasibility tests consistent with those of QP solver.