Atendendo a algumas respostas apreciativas e inúmeras ignorativas, aqui está......

02 fevereiro 2018

Derivative-free optimization

I uploaded the preprint "The Barycenter Method for Direct Optimization" onto arXiv. One thing I didn't mention is Bayesian optimization, a form of direct optimization which tries to learn about the shape of the function being optimized using statistical priors.

No mentioned is "Grad student descent", a term I have just learned, which seems to be due to @ryan_p_adams, although the method itself is as ancient as walking forwards. It may be better than my own barycenter method if grad student time is cheap enough.

Back to being serious, I am pretty sure that barycenter is better than Bayesian optimization, however hard working and poorly paid the grad students may be. The main difference from my point of view is that Bayesian optimization is trying to learn something about the complete function in the process of finding the minimum. That is bound to be more expensive and slow, plus it's sort of scary for a controls person - if you're using your candidate optimal to make decisions in real time, you want to be really careful exploring alternatives for the sake of knowledge.

The Bayesian optimization literature has ways to deal with such issues, but still - my guess is that in many applications the barycenter method is leaner, and more transparent from the point of view of implementing and understanding what goes one. Would welcome some testing - the necessary extensive simulations are something I don't have the grad student power to do.

Nenhum comentário: