roblanf / aiphylo Goto Github PK

One thing to discuss is whether or not we want to shuffle the sites of the alignment. In my opinion this would make it a better comparison, because it's a more biological way to represent heterotachy that is unaccounted for.

My concern here is that if we don't shuffle sites then any machine learning approach could, quite reasonably, learn to treat the first and second halves of the alignment differently. In this case, comparing machine learning to maximum likelihood might not be very meaningful - because we could easily set up a maximum likelihood analysis that also accounted for the difference between the two halves of the alignment. In essence, the machine learning would be learning to do better based on some fairly meaningless feature of our simulated datasets.

We can avoid this by shuffling the sites in the generated alignments, so that the sites from the two models do not occur in a predictable order. In this case, it's more reasonable to think that machine learning would be doing something meaningful if it can beat maximum likelihood.

Just a thought. Easy to implement.