Pipeline Complexity over time - How to approach a solution

Since TPOT allows lots of freedom for the solution spaces it searches, sometimes TPOT can explore unnecessarily complex pipelines.

## Possible fix

Last week I met with @weixuanfu and Trang to discuss how we could solve this problem by looking at Pipeline complexity. Based on our conversation we had the ideas below, I am interested in other's ideas.

**Ideas**

1) scoring pipeline complexity by number of features (of end features set vs original) and add complexity penalty
2) scoring pipeline complexity by the number of parameters in a pipeline and add a respective penalty
3) Provide a validation or test data set in addition to the train data set . Once the best model is scored using cross validation, that same model can be refit against the validation set. If the performance improves on the train set but drops in the validation set, there could be a complexity penalty put on that respective pipeline.