LINCOA instead of Nelder-Mead for fine optimization

avaneev / biteopt

Derivative-Free Global Optimization Method (C++, Python binding) - Continuous, Discrete, TSP, NLS, MINLP

MIT License

137 stars 9 forks source link

LINCOA instead of Nelder-Mead for fine optimization #8

Closed asenzz closed 3 weeks ago

asenzz commented 3 weeks ago

Hi, I just wanted to suggest you to try the LINCOA method or BOBYQUA instead of the current simplex-based search in BiteOpt. It may provide a better convergence rate. Both are available in the Prima package.

avaneev commented 3 weeks ago

What do you mean exactly? BiteOpt is not based on simplex search - the NMSeq is statistically used infrequently.

asenzz commented 3 weeks ago

At first glance it seemed that you are using NM to refine the solution. Thanks for this great piece of software btw.

avaneev commented 3 weeks ago

NM is only used as a "kick-starter" that supplies solutions sporadically. In that sense SpherOpt is much more effective, which has a higher priority. Beside that, NM implementation is different to usual "simplex" implementation - NMSeq was extended to support any population size - in this form it's quite a good global solver.