maschimax / hpo-benchmarking-framework Goto Github PK

https://github.com/maschimax/housing_regression/blob/1429622726cc0bb53851269b7ead3815335e74e9/hpo/robo_optimizer.py#L186
https://github.com/maschimax/housing_regression/blob/1429622726cc0bb53851269b7ead3815335e74e9/hpo/robo_optimizer.py#L189
https://github.com/maschimax/housing_regression/blob/1429622726cc0bb53851269b7ead3815335e74e9/hpo/robo_optimizer.py#L229
https://github.com/maschimax/housing_regression/blob/1429622726cc0bb53851269b7ead3815335e74e9/hpo/robo_optimizer.py#L232
https://github.com/maschimax/housing_regression/blob/1429622726cc0bb53851269b7ead3815335e74e9/hpo/robo_optimizer.py#L274
https://github.com/maschimax/housing_regression/blob/1429622726cc0bb53851269b7ead3815335e74e9/hpo/robo_optimizer.py#L277

https://github.com/maschimax/housing_regression/blob/1ed5ea03e63131fce8f2c364880470ac2ed4852d/hpo/robo_optimizer.py#L55

• Hpbandster / ConfigSpace samples outside of the defined boundaries for lightGBM's parameter 'bagging_fraction'
• BOHB samples values higher than 1.0, which causes an error

space_lgb = [skopt.space.Integer(2, 256, name='num_leaves'), skopt.space.Integer(20, 1000, name='min_data_in_leaf'), skopt.space.Integer(-1, 100, name='max_depth'), skopt.space.Real(low=0.3, high=1.0, name='feature_fraction'), skopt.space.Real(low=0.3, high=1.0, name='bagging_fraction')]

Error message: File "ConfigSpace/configuration_space.pyx", line 1378, in ConfigSpace.configuration_space.Configuration.init
ValueError: Trying to set illegal value '1.0073910188601063' (type '<class 'float'>') for hyperparameter 'bagging_fraction, Type: UniformFloat, Range: [0.3, 1.0], Default: 0.65' (default-value has type '<class 'numpy.float64'>')

• According to the scikit-optimize documentation, the default value for the parameter base_estimator in forest_minimize is 'ET', which refers to an extra tree regressor (https://scikit-optimize.github.io/stable/modules/generated/skopt.forest_minimize.html#skopt.forest_minimize)
• base_estimator must be set to 'RF' to use a random forest regressor as a surrogate model like in the original SMAC algorithm (https://ml.informatik.uni-freiburg.de/papers/11-LION5-SMAC.pdf)