Implementing Successive Halving
Similar to CFS, Successive Halving (SH) is also part of the Multi-Fidelity Optimization group (see Chapter 6). There are two implementations of SH in sklearn
, namely HalvingGridSearchCV
and HalvingRandomSearchCV
. As their names suggest, the former class is an implementation of SH that utilizes Grid Search in each of the SH iterations, while the latter utilizes Random Search.
By default, SH implementations in sklearn
use the number of samples, or n_samples, as the definition of the budget or resource in SH. However, it is also possible to define a budget with other definitions. For example, we can use n_estimators
in RF as the budget, instead of using the number of samples. It is worth noting that we cannot use n_estimators
, or any other hyperparameters, to define the budget if it is part of the hyperparameter space.
Both HalvingGridSearchCV
and HalvingRandomSearchCV
have similar standard SH parameters to control how the SH iterations will work...