Adjust LN value scaling to bring overrated LN maps more in line #24109
Conversation
Spent a couple days discussing this on the pp rework server about values that were the most acceptable, these seemed to be the best from the community standpoint of top players. Note: This is more to fix issues with the current system, not to be a final solution. Related Google Sheets Page: https://docs.google.com/spreadsheets/d/1P0AxfdKvMHwWBQder4ZkFGO1fC9eADSGCryA5-UGriU/edit?usp=sharing
bdach
changed the title
|
While I'm not qualified to review this change on merit, I've taken the liberty of editing the title of the PR to omit the problematic beatmap's name. While I understand this change may be intended to fix that one case in particular, I'd like to think that we don't hotfix specific maps but rather under- or overrated patterns, and the spreadsheet even shows that (other maps are also affected). So to avoid overfocusing on this one beatmap I think it's better to focus on the bigger problem which is overrated LNs. |
|
That's fine then, my bad. |
|
SR Changes are as follows: Top 20 Changes:
See the full table here: https://gist.github.com/Eve-ning/ac148fcabd6ed62badb806f630591cf3
|
|
Converted to draft whilst we look at implementing a small but crucial fix to how LNs are calculated |
Related Issue: ppy#22756 The trigger in question happens when (1) in a chord: a longer LN, then a shorter LN is processed respectively. (2) in a chord: a long LN, then a note is processed respectively. however, given the opposite processing step, it will fail to trigger. We observe that both situations have the same pattern, however has undeterministic results, which only depends on the order the mapper placed each note.
|
Top 250 most affected maps (Aug 2023 Dataset) https://gist.github.com/Eve-ning/966eac5e47f844e55fdfd5e6a16fc6fd |
|
changes look about what we expect from this so far, so should be good with the fixes in place |
|
I've added an SR/PP sheet for this in the OP. Please check it to make sure the values are as expected :) |
|
reviewed by me and few other people ; the changes are looking like pretty much as we hoped for them to do, a couple of normal maps do get hit in the crossfire and some are unfortunately a bit underrated now, but those are minor in comparison to the massive change it makes to overrated maps (now bringing them a lot more in line). from some further investigation as well this seems to bring people's top 50 scores closer by ~14-34% which is nice to see. overall, approved. |
|
nerfing ln pp but refusing to rank more rice maps that are actually worth something, wild |
We're working on fixing everything else, be patient. |
One of the biggest issues that came out of the last fix made towards bringing LNs more in line was that the scaling was too harsh on the formula, causing, as the biggest ranked example, Duplicity Shade to fall from 12.04 on nomod to only 11.41 stars on halftime, meaning that an SS on HT was worth the same as ~98.7% on NM.
After some talks with the community about what scaling works best, we finally settled on this smoother formula, also increasing the release_threshold to bring LN maps that are still severely overrated more in line again.
This will affect only LN heavy maps, mainly in the 6+* range, and we've pulled a list of 6+* maps and ran them through NM/DT/HT tests as shown here to show the extent of the SR changes being applied.
Duplicity Shade in this example goes from 11.58 NM -> 9.85 HT, a 1.73* difference, up from 0.63*, meaning an SS now matches ~94% acc on a NM run.
The other maps affected in this list are well-known, heavily overrated maps due to their LN abuse. Rice-style maps and actually difficult LN maps are essentially unaffected by this change.
SR/PP sheet: https://docs.google.com/spreadsheets/d/1zRlycntHiptir4GtEE5TTqsspTqR1vqVpaSjFS9EVW4/edit#gid=1700775385
As of 7d8cef8