naive random sampling pfront estimation

[R-M-Lee]

Motivation

I think it would be nice to have a very simple example showing how to estimate the Pareto front from a strategy containing trained surrogate models (that you don't want to change any more)

Have you read the Contributing Guidelines on pull requests?

yes

Test Plan

N/A

[jduerholt]

@R-M-Lee : Looks good to me! Thanks. Would it make sense to also add pymoo to it? But this can also be done later.

What would be your typical usecase for it?

You have run a bayesopt campaing and have distributed points around the front and now you want to estimate the front more accurately, in order to navigate it?

[jduerholt]

@R-M-Lee: Any updates on this PR?

[jduerholt]

@bertiqwerty @R-M-Lee @KappatC @LukasHebing

Hi all,

we discussed pareto fron sampling on the hackathon on wednesday and I find it a super nice feature, especially, when we would combine it with more sophisticated sampling using NSGA-2 and pymoo (since pymoo is now also a dependency). For implementing this we could reuse a lot of the code from @LukasHebing that he wrote for optimizing an acqf with pymoo, since from setting up pymoo out of the domain, it makes no diference if we optimize an acqf or surrogates. This could all stay ;)

I also included @KappatC since in the VBLL paper, they showed an interesting acqf for multiobjective opt based on thompson sampling and NSGAII. Also there we could then reuse the code. But first would be to have the VBLL surrogate.

Best,

Johannes

[LukasHebing]

@R-M-Lee: Did you check out the GA (mostly in bofire/strategies/predictives/utils.py)?
My idea is, that we could make most of the methods a bit more abstract. We only need to replace the acqf-function part with something more general