- Blind mode tutorial
lichess.org
Donate

Rating System Broken.

#31

Response to post #30:

That's all fair, but observe that my winrate against stronger opponents (1500-1600) is also about 40%. I used to resign games due to being forced to stop playing in the middle of a game, and I also played lots of piece sac gambits in rated games leading to losses as well.

Response to post #30: That's all fair, but observe that my winrate against stronger opponents (1500-1600) is also about 40%. I used to resign games due to being forced to stop playing in the middle of a game, and I also played lots of piece sac gambits in rated games leading to losses as well.
#32

@EnderShadow8 your problem has nothing to do with rating system but with the fact that you are just too strong. Not many +2000 players here. it also breaks the rating system at the high end. Not enough games with small enough rating differences at the high end. all formulas for expected outcome tend to be in accurate when skill difference goes high enough

@EnderShadow8 your problem has nothing to do with rating system but with the fact that you are just too strong. Not many +2000 players here. it also breaks the rating system at the high end. Not enough games with small enough rating differences at the high end. all formulas for expected outcome tend to be in accurate when skill difference goes high enough
#33

Excuse me? I'm only an average 1500-1600

Excuse me? I'm only an average 1500-1600
#35

Elo doesn't have any of this madness. I propose to switch to Elo

Elo doesn't have any of this madness. I propose to switch to Elo
#37

there is no need to dissect Glicko-2 there is abundant amount analysis of it and it way better than Elo which really dated and bad rating system. G2 is lot better and can be shown to be better. It is not great there are better ones but reasonable and takes very little computational resources. There is reason why Lichess switched to G2 algorithm. I could not found the blog entry as it quite old decision. And favored by everyone Chess.com uses glicko, FICS uses Glicko.

there is no need to dissect Glicko-2 there is abundant amount analysis of it and it way better than Elo which really dated and bad rating system. G2 is lot better and can be shown to be better. It is not great there are better ones but reasonable and takes very little computational resources. There is reason why Lichess switched to G2 algorithm. I could not found the blog entry as it quite old decision. And favored by everyone Chess.com uses glicko, FICS uses Glicko.
#38

https://lichess.org/forum/lichess-feedback/elo-rating-implementation#3
https://lichess.org/faq#ratings

Well, I have a big laugh, anyway. Particularly on this:

"When you first start using the site, your rating starts at 1500 ± 700. The 1500 represents your rating, and the 700 represents the confidence interval. Basically, the system is 90% sure that your rating is somewhere between 800 and 2200.
...
This allows you to gain/lose points points more rapidly to match any changes in your skill level over that time."

And this: "Elo is a special case of Glicko-2 which ignores the fact that players learn over time."

Probably I will open new experimental account where I will lose on purpose all first matches and start with stabilized 800, playing all further games with 1200-1900 rated opponents (400+ difference at least), thus we will see how this G2 implementation here works...

Would I will have then provisional 1500, more or less, assuming winning the same amout of games I lost previously?

I believe I will be flagged immediately for artificial decreasing/increasing rating or using computer assistance, no matter how fast I would play blitz games only and win all I lost previously with the same account.

https://lichess.org/forum/lichess-feedback/elo-rating-implementation#3 https://lichess.org/faq#ratings Well, I have a big laugh, anyway. Particularly on this: "When you first start using the site, your rating starts at 1500 ± 700. The 1500 represents your rating, and the 700 represents the confidence interval. Basically, the system is 90% sure that your rating is somewhere between 800 and 2200. ... This allows you to gain/lose points points more rapidly to match any changes in your skill level over that time." And this: "Elo is a special case of Glicko-2 which ignores the fact that players learn over time." Probably I will open new experimental account where I will lose on purpose all first matches and start with stabilized 800, playing all further games with 1200-1900 rated opponents (400+ difference at least), thus we will see how this G2 implementation here works... Would I will have then provisional 1500, more or less, assuming winning the same amout of games I lost previously? I believe I will be flagged immediately for artificial decreasing/increasing rating or using computer assistance, no matter how fast I would play blitz games only and win all I lost previously with the same account.
#39

That's an interesting thought. It would definitely prove the effectiveness of Glicko and hopefully my suspicions.

That's an interesting thought. It would definitely prove the effectiveness of Glicko and hopefully my suspicions.
#40

to prove it effectivness is easiest in simulation one sample is not enough as any measurement has noise. But yes if win or lose a lot in a row your RD will grow and you will gain lose more points per game. Same will happen if you do not play for a long time. Also there is derivate part of G2 that will detect that you are movin up or down constantly and speedup the process.

to compare result you can use FIDEand deloitte rating algorithm competition to get way better comparison compared to just one person trying:
from https://en.chessbase.com/post/can-you-out-predict-elo-competition-update

"Currently, out of 162 teams, the benchmarks hold the following rankings:

Chessmetrics Benchmark: #10
Glicko-2 Benchmark: #38
Glicko Benchmark: #39
PCA Benchmark: #66
Elo Benchmark: #82
"

Obviously none old fashionen algorithm hold against maximum likely hood algorithms. But difference is pretty clear in favor of Glicko vs Elo .
Most of Go sites use ML based algorithms. BayesElo is one if wanna see how they work. computational effort is at least 100 fold

to prove it effectivness is easiest in simulation one sample is not enough as any measurement has noise. But yes if win or lose a lot in a row your RD will grow and you will gain lose more points per game. Same will happen if you do not play for a long time. Also there is derivate part of G2 that will detect that you are movin up or down constantly and speedup the process. to compare result you can use FIDEand deloitte rating algorithm competition to get way better comparison compared to just one person trying: from https://en.chessbase.com/post/can-you-out-predict-elo-competition-update "Currently, out of 162 teams, the benchmarks hold the following rankings: Chessmetrics Benchmark: #10 Glicko-2 Benchmark: #38 Glicko Benchmark: #39 PCA Benchmark: #66 Elo Benchmark: #82 " Obviously none old fashionen algorithm hold against maximum likely hood algorithms. But difference is pretty clear in favor of Glicko vs Elo . Most of Go sites use ML based algorithms. BayesElo is one if wanna see how they work. computational effort is at least 100 fold

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