[FEATURE REQUEST]: Implementing Region Averaged Acquisition Functions

[SaiAakash]

Motivation

A very recent paper published in AAAI conference on AI (2025) talks about region-averaged acqusition functions that can perform efficient trust region selection for high-dimensional and multi-modal Bayesian Optimization problems.

Paper: Regional Expected Improvement for Efficient Trust Region Selection in High-Dimensional Bayesian Optimization

Code: Regional-Expected-Improvement

This improves upon the existing TuRBO method by picking trust regions efficiently during initializations and during local optimization restarts using region-averaged acquisition functions instead of randomly sampling trust regions. The code already uses BoTorch infrastructure to construct the REI, qREI, LogREI and qLogREI acqusition functions. This might be a nice feature to add to the optimization suite present in BoTorch.

Any comments on the feasibility of integration, suggestions are welcome !

Describe the solution you'd like to see implemented in BoTorch.

Probably a new script in the acquisition folder that implements all 4 of the region-averaged acquisition functions directly. The usefulness can be demonstrated through another tutorial for TuRBO with efficient trust region selection.

Describe any alternatives you've considered to the above solution.

NA

Is this related to an existing issue in BoTorch or another repository? If so please include links to those Issues here.

NA

Pull Request

I am willing to open a PR if everyone thinks this is something worth adding to BoTorch !

Code of Conduct

• I agree to follow BoTorch's Code of Conduct

[Balandat]

I've only skimmed the paper, but this seems simple enough that including it would be reasonable assuming this actually provides meaningful performance improvements. @dme65, @saitcakmak what do you think about the approach?

We could also first include this in botorch_community if we're not sure about the performance value-add.

[saitcakmak]

I'd be open to adding them to botorch_community. If we see meaningful improvements, we can then promote one of the methods (I'd guess this would be qLogREI, due to log related improvements and generality that comes from q-batch support) to core botorch.

[SaiAakash]

So, should I go forward an add these in the botorch_community folder ?