How do use Hyperopt.jl's Bayesian Hyperparameter Optimization feature?

That’s correct. The old GPSampler has been removed in a recent commit, and a new BOHB sampler will be added by merging

once I’m back in office.

The GPSampler was removed since it never worked very well. Bayesian optimization itself requires a lot of carefully tuned hyperparameters to work, and Hyperopt tried to pick some clever defaults. In practice it was very brittle and hard to understand why it did or didn’t work. It also required a lot of effort to maintain.

A new release with release notes will be out once all BOHB has landed and I’ve played around a bit with it.

2 Likes