What PEPs are appropriate for a given domain? Using the current architecture we can get a very crude guide to that information by considering the features selected at the feature selection and classification stages. However, it doesn't really use all of the data.
Perhaps it is possible to have a library of PEPs available and by analysing the training streams to determine which of the PEPs are appropriate. It might be considered how well each PEP models the data.
Further down the line, it may also be possible to consider techniques for looking at a domain and extracting the PEPs for that domain from training data.