LbfgsbSolver not working properly. Always return initial value. about cppnumericalsolvers HOT 6 CLOSED

The L-BFGS-B is very sensitive to the objective function and initialization of the Rosenbrock function. I can solve a dual SVM with the implementation. I know that the step-width function is not the best choice yet. Thats on my todo-list.

from cppnumericalsolvers.

Have you made any progress on this issue? I am having the exact same problem. Lbfgs does make progress, but Lbfgsb only returns the initial guess.

Interestingly, if I add print statements inside my objective function, I can see that it is evaluating points with a better residual, but doesn't give them back?

from cppnumericalsolvers.

There was anther paper about L-BFGS-B I was not aware of. Before making any new commit I have to debug my implementation of the HZ update rule which would find points satisfying both Wolfe-conditions. That's definitely the plan: making L-BFGS-B working on the Rosenbrock function.

from cppnumericalsolvers.

Okay - I'm looking forward to trying it when you have finished debugging!

from cppnumericalsolvers.

I added the linear search of MoreThuente. However, the refactoring destroyed the L-BFGS-B.

from cppnumericalsolvers.

See unittest. Lbfgsb is mow working again

from cppnumericalsolvers.