question about Implementing Turbo-m

[Yoonho-Na]

https://botorch.org/docs/tutorials/turbo_1/

Hi I'm currently doing experiment on turbo method.
I did turbo-1 using above link but it doesn't provide explanation about Turbo-m (such as Turbo-20)
How to implement turbo-m ?
If there is an example code for it?

[saitcakmak]

Hi @Yoonho-Na. We don't have an example for Turbo-m. The general idea is very simple though and it can take various forms depending on how closely you want to follow the original paper. You'd need to keep multiple trust regions and track all their centers as part of the Turbo state. You'd need a heuristic to determine how to select the center points for each TR (such as best point with a minimum distance to other centers), since you'll likely not want the trust regions to overlap too much. You can then determine how to split the candidate generation between each trust region. You can generate candidates from each one and just pick the "best", or distribute a batch of candidates between different trust regions using a heuristic (such as generating from each TR and picking one with probability proportional to the acquisition value).