EmuKit/emukit

problem concerning convergence

AmosJoseph opened this issue · 3 comments

Hi!

I try expotential and Matern 5/2 kernal for 10 variables,but do not converge or with a trend to converge after 40 iterations.
Could you give any suggestion?

How about building (sparse) GP with evaluated samples to run iteration for BO?
Will this converge?

Thank you!

Best!

That is very highly dependent on the problem. Such behavior can be happening for a ton of reasons. Maybe the objective is all bumpy, with many local minima, maybe there is no structure in the parameter space, maybe the objective is noisy. I am afraid I don't have any generic advice to give here.

What do you mean by "maybe there is no structure in the parameter space"?

That's more about the cases when you have many categorical variables. There are techniques that could be used in such cases: http://proceedings.mlr.press/v80/lu18c.html . But if the space is unstructured, you optimization essentially becomes brute force no matter what you do