It's easier said than done, but you might want to approach this like a data science project. Record the data from your test runs, annotate the behaviour you want, and build up a dataset. You should be able to experiment with any number of tweaks (admittedly except for the data sampling itself) without having to get back on the bike.