Problem when Sobol falls back to HitAndRunPolytopeSampler about ax HOT 5 OPEN

Thanks for flagging this, this does seem like a bug. I'll take a closer look

from ax.

OK so the issue is that the hit and run sampling fallback in Ax (see here currently neither applies the normalization that you mentioned nor, and more importantly, does it apply thinning to the samples, which results in much more correlated samples. Basically, we need to update the Ax logic to call get_polytope_samples instead. This should be straightforward, the only thing to be careful about is how to manage the handling of the seed between draws so things can be made to work deterministically. I can put up a PR for this.

from ax.

Thanks for looking into this. Sounds right, may as well call it directly since it seems to be working fine. When you say thinning what exactly does that mean? Sorry not 100% familiar with all the jargon yet.

from ax.

Thinning (removing all but every k sample) is a process to reduce autocorrelation of samples drawn from a MCMC chain. This is happening here: https://github.com/pytorch/botorch/blob/main/botorch/utils/sampling.py#L828 (this is a poor man's implementation, you could avoid storing the discarded samples as you go to save memory but this is typically not of concern for the sample sizes we use for Bayesian Optimization.

from ax.

Got it, thanks.

from ax.