Seems like the behavior becomes predictable and then you have to retrain if you see unoptimal results.