smac.facade.smac_bb_facade¶

Classes

SMAC4BB([model_type])

Facade to use SMAC for Black-Box optimization using a GP.

class smac.facade.smac_bb_facade.SMAC4BB(model_type='gp_mcmc', **kwargs)[source]¶

Bases: smac.facade.smac_ac_facade.SMAC4AC

Facade to use SMAC for Black-Box optimization using a GP.

see smac.facade.smac_Facade for API This facade overwrites options available via the SMAC facade

Hyperparameters are chosen according to the best configuration for Gaussian process maximum likelihood found in “Towards Assessing the Impact of Bayesian Optimization’s Own Hyperparameters” by Lindauer et al., presented at the DSO workshop 2019 (https://arxiv.org/abs/1908.06674).

Changes are:

Instead of having an initial design of size 10*D as suggested by Jones et al. 1998 (actually, they suggested 10*D+1), we use an initial design of 8*D.
More restrictive lower and upper bounds on the length scale for the Matern and Hamming Kernel than the ones suggested by Klein et al. 2017 in the RoBO package. In practice, they are np.exp(-6.754111155189306) instead of np.exp(-10) for the lower bound and np.exp(0.0858637988771976) instead of np.exp(2) for the upper bound.
The initial design is set to be a Sobol grid
The random fraction is set to 0.08447232371720552, it was 0.0 before.

SMAC3 Documentation

smac.facade.smac_bb_facade¶