I mean, in handling boxed constraints?

In terms of stability, and more importantly, the numerical performance?

I have already written some well-optimized and well-tested C/CUDA/C++ codes for several unconstrained optimzation methods.

And I use augmented largrangian to handle constraints, however, it looks like bounded BFGS is quite different than the standard BFGS, I cannot do it over my existing BFGS routine,it require almost a complete re-written, I dont really want to do that unless there is significant gain there.

Have anyone, with practical experiences, can tell me if the rewritten is worth doing or I can stick to the let Lagrangian to handle it?

Btw, I usually use BFGS to solve MLE type optimzation problems, where the evaluation of the objective function is costly.

  • $\begingroup$ Did you ever find a solution to implementing bounded bfgs on the GPU? I'm currently trying to implement it in Tensorflow. $\endgroup$ – mdornfe1 Jun 18 '17 at 19:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.