Mismatched keywords between TPOT and scikit-learn

[nathanWolo]

When using verbose=3 on the TPOTRegressor many warnings are given about mismatched keyword arguments, such as GradientBoostingRegressor being passed 'ls' instead of 'squared_error' as the metric during the mutation step.

Examples:

_pre_test decorator: _mate_operator: num_test=0 'str' object has no attribute 'arity'.

_pre_test decorator: _random_mutation_operator: num_test=0 The 'loss' parameter of GradientBoostingRegressor must be a str among {'huber', 'quantile', 'squared_error', 'absolute_error'}. Got 'ls' instead..

_pre_test decorator: _random_mutation_operator: num_test=0 The 'loss' parameter of SGDRegressor must be a str among {'huber', 'squared_epsilon_insensitive', 'epsilon_insensitive', 'squared_error'}. Got 'squared_loss' instead..

_pre_test decorator: _random_mutation_operator: num_test=0 The 'loss' parameter of GradientBoostingRegressor must be a str among {'huber', 'quantile', 'squared_error', 'absolute_error'}. Got 'lad' instead..

Process to reproduce the issue

Fit TPOTRegressor with the newest versions of TPOT and scikit-learn.

Possible fix

Update the kwargs to match what is currently used in scikit-learn (or find an alternative to hard-coding them as string literals)

[perib]

The 'arity' error is due to a bug in the TPOT mutation function. I implemented a fix in this pull request #1268

The other errors are due to invalid hyperparameters being included in the config dict found here: https://github.com/EpistasisLab/tpot/blob/master/tpot/config/regressor.py

Should be as easy as just removing the invalid hyperparameters from the list.

@nickotto