Use hetGP for fitting heteroskedastic noise

[robsmith11]

I attempted to use mlrMBO with AEI for optimizing some very noisy (1-to-10 signal-to-noise) simulation data, but it fails to explore large regions of the domain because I believe the noise estimates are far too low.

I think that explicitly modeling the noise by replication of the noisy objective function for the same parameters could greatly help. The hetGP package seems to do this in an efficient way, but is lacking optimization tooling -- it appears to just focus on fitting the GP.

I'm planning to either write some basic optimization functions around hetGP or to figure out how to integrate it with mlrMBO. I'm not familiar with the internals of either package, so looking for any suggestions on the best approach. Is this a feature that would be appropriate to add to mlrMBO?

[ja-thomas]

Hi,

Any mlr learner can be used by mlrMBO as surrogate model, so in general hetGP needs to be integrated as an mlr learner, which shouldn't be too difficult. (see documentation here: https://mlr-org.github.io/mlr/articles/tutorial/create_learner.html)

You might also be interested in PR #430, which implements quite a few new noise handling strategies, but is not quite done yet, but you should be able to use the branch nevertheless.

[jakob-r]

It would be very interesting if you wrapped this learner in mlr. This way we can do benchmarks much easier. Let us know if you are willing to go that way. Then we could support you a little.