Okay, here are a few of my thoughts on the current system.
First off, yes, accuracy seems to be slightly overrated atm - but the much bigger problem lies in the as-of-yet mediocre difficulty calculation.
Just to prove that point I decided to quickly play three of the easiest maps from my top performance list to show how simple it is to beat them:
If you look at the time stamps, you can tell that these were
all first try (well, you'll have to trust me on the first score). These are now
#1,
#3 and
#5 in my overall top performance list. None of those scores have outstanding accuracy either (compared to my usual scores), so the huge pp gain must stem from the difficulty assesment.
For Against Gilgamesh I can at least somewhat understand why it's rated so highly, it has several comparatively long streams and two 1/6 burst sections.
Mythologia's End is more confusing since it consists almost entirely of basic patterns. The density is not amazingly high, neither is the BPM. It's not an "easy" map per se, but definitely not as hard as your algorithm makes it out to be.
Seikan Hikou is the most baffling of the bunch. I have absolutely no explanation why this has such a high pp yield. Being a HR DT score you might suspect that accuracy is actually the issue here, but this map was in my top 3 performances before the last update. Back then, accuracy was not overrated AND my accuracy on the map was (significantly) worse than it is now. So it's definitely some other factor... But the map is super basic, not a single long or tricky pattern. BPM and density are very reasonable as well. Oh, and this is currently my #1 performance by a huge margin of 19pp... Tom, could you maybe run this through your difficulty calculator and tell us which difficulty value is responsible so we can work on improving the algorithm?
Another example for weird difficulty behavior would be
Kakuzetsu Thanatos.
This map has two Oni level Taiko diffs. qoot's diff is way harder in my opinion; the patterns are significantly more difficult and the rhythm is a lot more complex. It has a few less objects, but that doesn't outweigh the other difficulty aspects at all (it does in fact feel faster to me when playing, despite the apparently lower density. Maybe 714's diff has the notes spread out more evenly?). The algorithm however rates 714's diff 0.10 stars harder, which is quite a big margin in this new system. Now I'm wondering if that's just because of the slightly higher object count - in that case, the weightings for all difficulty values need to be reconsidered. Or does your algorithm actually rate 714's diff higher in pattern/rhythm complexity? In that case, we'd probably need to work on the specific algorithm.