Objective vs. PosteriorTransform

[jduerholt]

Hi all,

I want to understand in which situations one should use a posterior transform in a single objective acqf and in which an objective module. For me, the functionality looks very much the same, for example when one compares the ScalarizedPosteriorTransform to the LinearMCObjective. Any advise here?

Best,

Johannes

[saitcakmak]

Hi @jduerholt. The PosteriorTransforms were originally called something like AnalyticObjective, so they're indeed similar. A key difference is that the PosteriorTransform operates on the Posterior object before sampling from it, while the MCObjective operates on the samples drawn from the posterior. This makes it possible to use PosteriorTransform with analytic acquisition functions that don't rely on sampling. However, having to analytically transform the Posterior object limits their flexibility.

MCObjective is quite flexible since it just needs to accept a tensor and return a transformed tensor, so it can support arbitrary callables. The limitation is that before you can use it, you need to sample from the Posterior. For seemingly identical transforms like ScalarizedPosteriorTransform and LinearMCObjective, there's also a slight difference in the computations. ScalarizedPosteriorTransform constructs a new posterior object after modifying the mean and mvn, but the sampling from this transformed posterior is cheaper since it has fewer outcomes. LinearMCObjective does not require any modifications to the posterior, but it requires sampling from the full posterior and processing those samples after, which can be more expensive & require more memory. I haven't benchmarked it but if I had to pick between these two, I'd lean towards ScalarizedPosteriorTransform.

[jduerholt]

Hi @saitcakmak, thank you very much for this detailed answer. It made it really clear to me!

[eytan]

Curious if folks have any ideas on how we can improve our docs to make this more clear (and if anybody would like to volunteer to help : )

…

On Mon, Jan 8, 2024 at 2:37 PM Johannes P. Dürholt ***@***.***> wrote: Hi @saitcakmak <https://github.com/saitcakmak>, thank you very much for this detailed answer. It made it really clear to me! — Reply to this email directly, view it on GitHub <#2164 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAW34P7MEAADWFK3LIED7LYNRDHDAVCNFSM6AAAAABBNQZTLCVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4DANJUG43DK> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>