SMAC3: A Versatile Bayesian Optimization Package for Hyperparameter Optimization

[nabenabe0928]

SMAC3: A Versatile Bayesian Optimization Package for Hyperparameter Optimization

Main points

• Description of the new version of the SMAC (v1.2)
• there are four options SMAC4BB, SMAC4HPO, SMAC4MF, SMAC4AC
• SMAC4BB is for low-dimensional continuous BBO and it uses GP
• SMAC4HPO is for hierarchical search space and it uses RF
• SMAC4MF uses BO+HB and this BO relies on RF
• SMAC4AC is the very first proposition of SMAC

Performance over time

x-axis: runtime (log-scale, second) y-axis: loss

1. Dragonfly
2. BOHB from the original paper
3. Hyperband
4. random search
5. SMAC-HB
6. SMAC (RF)

The target benchmarks are:

1. BNN in BOHB
2. HPOlib
3. NB1shot1