Help in resolving BadInitialCandidatesWarning

[shikhar-saxena]

I am getting the warning while using qExpectedImprovement Acquisition Function (for q=1)

BadInitialCandidatesWarning: Unable to find non-zero acquisition function values - initial conditions are being selected randomly

Can someone tell me how to resolve this? I increased num_restarts (from 5 to 10) and raw_samples (from 512 to 1024) but that didn't help.

[saitcakmak]

How does your best_f value compare with train_Y? This warning would happen if the acquisition function is unable to generate posterior samples that are better than best_f.

[shikhar-saxena]

I'm taking best_f = train_Y.max().item().

This warning would happen if the acquisition function is unable to generate posterior samples that are better than best_f.

So basically, my best_f value is too high for the algorithm to find some improvement with respect to it. Not sure how I can fix this. What does botorch do in this case btw?

I'm also using PI acquisition function (for this same setting). And PI doesn't give me this warning.

[Balandat]

This seems like a posterchild use case for the new qLogEi, which addresses these issues and is about to be merged in: #1936

[shikhar-saxena]

Thanks. I tried running my code with LogExpectedImprovement, and it no longer gives the warning.

Can you please let me know when qLogExpectedImprovement will be released?

[Balandat]

Great to hear. It'll be in the next release. We don't have a specific schedule for this one - @sdaulton were you planning to get other things into a release before next week?

[jakobzeitler]

Thanks so much, very helpful! Using this right now.