Does anyone know the TrueSkill configuration values (see
https://github.com/atomflunder/TrueSkill-Calculator?tab=readme-ov-file#configuration for a description) used by QM?
To add more color to this for MMR nerds:
above 500, they're straight TrueSkill and designed to give a fair competition.
Afaiu, TrueSkill is sensitive to beta and especially tau, so there isn't one "straight" TrueSkill (you can get some weird rating systems by selecting inappropriate values for the setting constants). QM further alters TrueSkill by:
1) not letting ratings (mu's?) change by more than 10 or less than 1
2) below a certain rating (I think 500), halving rating decreases
I think there's still a heavy activity component to Quickmatch ratings because (2) causes persistent rating inflation while (1) causes very good players to gain more than they should from lopsided matches.
For an example of (1), suppose you're a very good player playing RT games on a less popular template. You can keep constantly facing players you should win against 95% of the time based on your ratings. Without the min(10, max(1, calculated rating change)) limitation, your rating stays stable if you just win 95% of the time, goes up if you win more often, and goes down if you win only, say 92% of the time. In reality, your rating will go up as long as you win 11 times for every loss, so the ratings never stabilize at the very top unless the matchmaker can limit lopsided matchups or the template is one where it's very hard to win >91.7% of the time (like Small Earth Auto Dist) no matter the skill difference.
(2) is hard to think about because Quickmatch isn't like Elo where the average rating stays constant and you just have symmetric exchanges (winner gains x points, loser loses x points) but instead has a mean (mu) and standard deviation (sigma) for each player's rating estimate; that said, I believe it still behaves like a symmetric exchange system where the average rating should stay constant.
In theory, if you had a stable pool of players that just played one another constantly, Quickmatch ratings would converge to "straight" TrueSkill. In practice, players leave (with far-from-converged ratings), join (inflating the average rating through asymmetric exchanges), and play in ways that make it hard to matchmake (which can lead to lopsided matchups).
I don't think your rating ever
converges to your "true" rating when there's persistent inflation, but if you look at, say, the SEAD leaderboard on QM over a month, you will see that for the top active players the system has sort of figured out their real skill. It's pretty YMMV
Edited 8/17/2024 22:13:33