forum.alglib.net

Dear ALGLIB team and community,

I'm writing a paper on nonconvex optimization and I'm comparing several state-of-the-art solvers. I couldn't figure out how the ALGLIB SQP method is implemented, specifically:
- how it measures progress (filter or merit function). I think it then uses a line search to find an appropriate step length;
- how it deals with infeasible QPs. Does it relax the RHS or revert to minimizing constraint violation?
Hope you can help me out! Thanks a lot,

Charlie Vanaret

_________________
Check out my next-generation solver for nonconvex optimization: https://github.com/cvanaret/Uno

Hi!

ALGLIB uses an L1-penalized merit function to measure progress. It is a trust region method, i.e. it does not use a line search.

It uses SL1QP approach towards constraints and does not have separate feasibility enforcement stage. The merit function coefficient is dynamically chosen to ensure that there is a progress towards minimizing constraint violation.

Sergei.

Hi Sergey,
Perfect, thanks for your answer. I'll post a link to the preprint once our article is submitted :)
Best,
Charlie

_________________
Check out my next-generation solver for nonconvex optimization: https://github.com/cvanaret/Uno

Actually, it is planned to switch to filters at some moment later. They seem to be more promising. But still the filter version will inherit many features of the original SL1QP one.