-
|
There is a discussion related to how the TPE sampler of optuna handles dynamic changes of the search space at https://optuna.readthedocs.io/en/stable/faq.html#what-happens-when-i-dynamically-alter-a-search-space and more precisely at optuna/optuna#822. As I was planning to use this functionality from the BoTorchSampler as well, I was wondering how it was internally handled? Technically speaking it does not complain but the improvements I have observed so far are not much better than random search, which makes me qustion whether I have been using this library in an unintended way? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
When you say "BoTorchSampler", what exactly do you mean? The Optuna BoTorch itself, being a lower-level library, does not actually have the concept of a search space or a study. So from the point of view of BoTorch, this kind of thing would just mean that you define new |
Beta Was this translation helpful? Give feedback.
-
|
Thank you @Balandat! |
Beta Was this translation helpful? Give feedback.
When you say "BoTorchSampler", what exactly do you mean? The Optuna
BoTorchSampler? I am not familiar with the implementation of that.BoTorch itself, being a lower-level library, does not actually have the concept of a search space or a study. So from the point of view of BoTorch, this kind of thing would just mean that you define new
bounds(or a set of new discrete inputs if you're doing discrete of mixed optimization), fit a new model and optimize the acquisition function on whatever that new domain is while including the data from the previous evaluations.