[Q] Differences between SMAC4MF and HpBandSter?

[mhlr]

Both SMAC4MF and the HpBandSter package are implementations of the BOHB algorithms by the same github org. HpBandSter does not appear to be actively developed.

What are the differences between these implementations?

What are potential reasons to use HpBandSter over SMAC4?MF?

[mlindauer]

Hi, HpBandSter is the original implementation of BOHB. However, the HpBandSter repo is no longer actively maintained by us after the main developer left the group. Therefore, we reimplemented the BOHB idea in SMAC. Basically, it is the same, but SMAC does not implement the KDE models, but supports GPs and RFs as surrogate models. Our recent studies for HPOBench showed that SMAC4MF with RFs often outperforms HpBandSter with KDEs.

[NLGithubWP]

In HpBandSter, there is framework for tuning the parameter in parallel with master, dispatcher, and multiple workers, But in SMAC, is there similar framework for running BOHB in parallell ?

[alexandertornede]

Yes, SMAC 2.0 supports running it in parallel. An example on how to use multiple workers can be found here: https://automl.github.io/SMAC3/v2.0.0/3_getting_started.html#scenario