lichess.org
Donate

Suggestion: Lichess Rating System improvement (Glicko1 > Glicko2 for live chess servers)

The current rating system (Glicko-2) works reasonably well here, but as you might notice, if you go to players' rating profiles, they never have an RD less than 60. The formulas of Glicko-2 (specifically volatility related formulas) prevent RD from decreasing under 60 or roughly thereabouts when the rating updates after a single game result.

Why does this happen? It's complicated but in the words of the author of the Glicko-2 system: "The Glicko-2 system works best when the number of games in a rating period is moderate to large, say an average of at least 10-15 games per player in a rating period". It has to do with the way RD' is calculated in the Glicko-2 system and because it's a function of the new variable volatility, which is designed in the Glicko-2 system to update after a moderate amount of games at a time (10-15+) rather than a single game, as single games do not allow the formulas needed to calculate volatility enough data to accurately measure how far a single game performance should vary from the expectation per rating update period. Volatility is absent in the regular Glicko system.

However RD should keep decreasing until a floor of about 40 if a player is very active as their rating should become more statistically significant if they play many games in a short period. A minimum RD of ~60 simply does not reflect this. The regular Glicko system allows RD to keep decreasing to the recommended floor of 30-50, rather than 60.

Additionally, the formulas for the regular Glicko system, which beyond being much more simple than Glicko-2, also allow RD to increase as a function of time (periods of inactivity). Which would also be a more efficient solution for choosing which players should be shown on leader boards or have active ranks (could simply be a condition of RD < 100 is required to be on the leader board for example)

Tl;dr: The Glicko rating system is better than Glicko-2 for live servers where ratings update after every game played as Glicko-2 was designed to work for ratings to update after a medium to large (10-15+) sample of games. Implementing Glicko would solve the issue of RD not decreasing under 60 (which it definitely should) and also allow players' RD to increase after periods of inactivity.
Interesting... I tried to verify your tl;dr and found: chess.stackexchange.com/a/1261

Glickman writes about Glicko http://glicko.net/glicko/glicko.pdf :
"The Glicko system works best when the number of games in a rating period is moderate, say an average of 5-10 games per player in a rating period. The length of time for a rating period is at the discretion of the administrator."

I think what this means is that Elo or some other system would be more accurate than either Glicko or Glicko-2, if Lichess is to violate the assumptions stated on the first couple pages of either (Glicko or Glicko-2) PDF (that games played in a rating period are rated concurrently insofar as ordering is concerned; and the recommended number of games per period).
Yes, Glicko1 was designed for more than one game at a time as well, but the important thing about Glicko1 is that it still allows RD to keep decreasing to a specified floor after each single game played (evidence of this can be seen on rating stats of active players of chess.com, where they use Glicko1) whereas Glicko2 does not allow that. Elo makes too many assumptions to be better than Glicko1, it relies only on a single outcome probability formula to update ratings and doesn't factor in rating accuracy.
Because there is an optimum number (5-10) of games in a rating period for the Glicko system, it seems to me our ratings shouldn't update after every game. Also, the duration of the rating period for Blitz/Bullet should be substantially less than the rating period for Correspondence.

Custom fitted rating periods and some way of dealing with large RDs would go a long way to solving my small issue with losing to a (1500?) who, 10 games later is a non-provisional 2200. Rating updates after every game = absence of rating period, no?

As for RD, leaderboards, active and inactive players, I'm with Dragonair.

Mike

This topic has been archived and can no longer be replied to.