How to optimize the acquisition function without gradient-based method

[MJeong19]

Currently, I am trying to make a code for new acquisition function.
However, an error occured in the line with optimize_acqf() because the output of the acquisition function did not have gradient.
So I would like to know if there is a way to optimize the acquisition function without requiring gradient,
such as random process.

[eytan]

Hi there! This tutorial should help you with this exact problem. https://botorch.org/docs/tutorials/optimize_with_cmaes/ May I ask more about your setup and why the AF doesn’t have a gradient? Are you using some sort of custom model? Best, E

…

On Thu, Jun 26, 2025 at 11:10 AM MJeong19 ***@***.***> wrote: Currently, I am trying to make a code for new acquisition function. However, an error occured in the line with optimize_acqf() because the output of the acquisition function did not have gradient. So I would like to know if there is a way to optimize the acquisition function without requiring gradient, such as random process. — Reply to this email directly, view it on GitHub <#2896>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAW34I5AS47ERQI66NQY2L3FQEOZAVCNFSM6AAAAACAGQU26SVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZYGQ4TQOJTHA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[hvarfner]

Hi @MJeong19 ,

You can also go with

candidate, acq_value = optimize_acqf(
    acq_function=your_acquisition_function,
    bounds=bounds,
    q=1,
    num_restarts=4,
    raw_samples=256,
    options={"with_grad": False}, <--- LOOK HERE
)

if you want another option. However, this may yield worse candidates (and be slower) than using gradients for optimization. If you can provide more info or a small repro, we'd be happy to help you figure out what's wrong!