Unstable models

I want to suggest an improvement for TPOT. After evaluating many pipelines on different regression datasets I came to the conclusion that some models are less suitable for AutoML as they return instable results: Running the model multiple times with different random seeds give completely different results (R2 vary between 0.4 and 0.6) which can lead to a good candidate during optimization by chance and bad performance on the test dataset. I experienced that behavior for examlpe with RandomForestRegressor/ExtraTreesRegressor (only with estimators > 1000 it stabilizes), LassoLarsCV and LinearSVR.

In practice I would manually tune these model until I get stable results, but I didn't find a generic parameter space that brings stable results for all datasets.

For the moment I had to remove those models from my TPOT configuration. As an suggestion for improvement models could be tested for stability by training and validating them with different seeds before they are added as candidates to the TPOT population.

What you think about it?