There are Lichess users that put FIDE rating on their profile. Based on this I get the users Blitz rating and do regression after removing unbelievable data points and outliers.
Does it change with time? 2 Years ago I had 1850 rating on lichess, and I have improved since then, and yet, my opponents seem to have gotten a lot tougher, and I can't make it past 1650. This seemed to coincide with the separation of blitz and rapid time controls, or something, so my rating was 200 elo higher before the separation, I wonder if this has some influence on the calculation (or if people have already adjusted and this doesn't matter, but then only people that have played at least 100 games since then should be taken into account.)
Your beliefs create your reality, so be careful what you wish for.
Run another regression. Outliers such as abs(lichess_blitz_rating - fiderating) > 400 are not included. Data are mostly from members of teams such as streamers-union, bengal-tiger, pawn-stars-2, crestbook-chess-club, zhigalko_sergei-fan-club, fide-checkmate-coronavirus, lichess-swiss and top players from blitz and rapid leaderboards.
Linear Regression
model : xgboost linear regressor
Total Datasets : 149368
Feature : ['blitzrating']
Target : fiderating
Min Blitz Games : 50
Datasets used in regression: 5727
Some top players by fide rating who are included in the regression. This provides better rating estimates for top players. The top is Magnus followed by Wesley. I am not sure if all of these players in this particular list have the correct FIDE rating.
Ferdy wrote: ↑Tue Dec 21, 2021 4:52 pm
Run another regression. Outliers such as abs(lichess_blitz_rating - fiderating) > 400 are not included. Data are mostly from members of teams such as streamers-union, bengal-tiger, pawn-stars-2, crestbook-chess-club, zhigalko_sergei-fan-club, fide-checkmate-coronavirus, lichess-swiss and top players from blitz and rapid leaderboards.
Linear Regression
model : xgboost linear regressor
Total Datasets : 149368
Feature : ['blitzrating']
Target : fiderating
Min Blitz Games : 50
Datasets used in regression: 5727
Some top players by fide rating who are included in the regression. This provides better rating estimates for top players. The top is Magnus followed by Wesley. I am not sure if all of these players in this particular list have the correct FIDE rating.
This is interesting and quite reasonable; the roughly 6 to 5 spread in ratings from FIDE to Lichess blitz is about what I would expect due to the smaller percentage of draws in blitz. Is there any data that would allow you to compare CCRL blitz ratings with Lichess blitz ratings? Due to the compression of CCRL ratings by BayesElo and the expansion of LiChess blitz ratings, I would expect that the relationship would be fairly close to linear.
Looking at my profile on lichess and fide, I can imagine some 150 overrated score at lichess compared to Fide. I play a lot of correspondence games but with bullet kind of response time in sub-optimal conditions (on phone while driving etc...) and lichess rating corresponds to my actual fide rating while playing way below my tournament level.
ydebilloez wrote: ↑Tue Dec 21, 2021 7:07 pm
Looking at my profile on lichess and fide, I can imagine some 150 overrated score at lichess compared to Fide. I play a lot of correspondence games but with bullet kind of response time in sub-optimal conditions (on phone while driving etc...) and lichess rating corresponds to my actual fide rating while playing way below my tournament level.
Curious - I have not played on lichess in a while but many people like those 'Arena tournaments' where it's team against team and the sole goal is to play as many games as you can and score as many points as you can...weighted by # of consecutive wins.
To that end, the games may officially be 10 minute games + 5 sec, but the (usually) higher rated player has his time cut in half to 5 min/sd while his opponent may be using the full 10 min. Beserking, I think they call it.
That could affect peoples ratings...but I guess if it's just a small percentage of games it would have minimal influence.
lkaufman wrote: ↑Tue Dec 21, 2021 5:32 pm
This is interesting and quite reasonable; the roughly 6 to 5 spread in ratings from FIDE to Lichess blitz is about what I would expect due to the smaller percentage of draws in blitz. Is there any data that would allow you to compare CCRL blitz ratings with Lichess blitz ratings? Due to the compression of CCRL ratings by BayesElo and the expansion of LiChess blitz ratings, I would expect that the relationship would be fairly close to linear.
There are maia bots in lichess like maia1 has around 1478 blitz. It is just lc0 + maia nn called maia-1100.pb.gz. This can be run with nodes=1 limit, you may use the older versions like Lc0 v0.26. Better to use CPU version to generate more games in shorter time if cpu has more threads.
maia9 has around 1770 blitz using maia-1900.pb.gz nn. This can be easily compared with CCRL engines around this strength levels.
There are other lichess bots but reproducing it for CCRL on our PC can be a challenge. maia is easier. I think the authors of those bots can run those under CCRL settings.
There is minihuman by dkappe with a higher blitz rating. Reading the description I think this can be easily adapted for CCRL. It has a limit of around 2500 nodes.
Mean Girl 8 (32x4) -- the most fun leela-style network -- looking at ~2500 nodes on a Raspberry Pi 3. Will play casual and rated Blitz, Rapid and Classical with at least 3 sec increment in both standard and chess960. Will move almost instantly. Designed to be a reasonable sparring partner. Uses a gambit book for extra entertainment value.
lkaufman wrote: ↑Tue Dec 21, 2021 5:32 pm
This is interesting and quite reasonable; the roughly 6 to 5 spread in ratings from FIDE to Lichess blitz is about what I would expect due to the smaller percentage of draws in blitz. Is there any data that would allow you to compare CCRL blitz ratings with Lichess blitz ratings? Due to the compression of CCRL ratings by BayesElo and the expansion of LiChess blitz ratings, I would expect that the relationship would be fairly close to linear.
There are maia bots in lichess like maia1 has around 1478 blitz. It is just lc0 + maia nn called maia-1100.pb.gz. This can be run with nodes=1 limit, you may use the older versions like Lc0 v0.26. Better to use CPU version to generate more games in shorter time if cpu has more threads.
maia9 has around 1770 blitz using maia-1900.pb.gz nn. This can be easily compared with CCRL engines around this strength levels.
There are other lichess bots but reproducing it for CCRL on our PC can be a challenge. maia is easier. I think the authors of those bots can run those under CCRL settings.
There is minihuman by dkappe with a higher blitz rating. Reading the description I think this can be easily adapted for CCRL. It has a limit of around 2500 nodes.
Mean Girl 8 (32x4) -- the most fun leela-style network -- looking at ~2500 nodes on a Raspberry Pi 3. Will play casual and rated Blitz, Rapid and Classical with at least 3 sec increment in both standard and chess960. Will move almost instantly. Designed to be a reasonable sparring partner. Uses a gambit book for extra entertainment value.
List of bots from lichess-bots team with 500 or more blitz games.
lkaufman wrote: ↑Tue Dec 21, 2021 5:32 pm
This is interesting and quite reasonable; the roughly 6 to 5 spread in ratings from FIDE to Lichess blitz is about what I would expect due to the smaller percentage of draws in blitz. Is there any data that would allow you to compare CCRL blitz ratings with Lichess blitz ratings? Due to the compression of CCRL ratings by BayesElo and the expansion of LiChess blitz ratings, I would expect that the relationship would be fairly close to linear.
There are maia bots in lichess like maia1 has around 1478 blitz. It is just lc0 + maia nn called maia-1100.pb.gz. This can be run with nodes=1 limit, you may use the older versions like Lc0 v0.26. Better to use CPU version to generate more games in shorter time if cpu has more threads.
maia9 has around 1770 blitz using maia-1900.pb.gz nn. This can be easily compared with CCRL engines around this strength levels.
There are other lichess bots but reproducing it for CCRL on our PC can be a challenge. maia is easier. I think the authors of those bots can run those under CCRL settings.
There is minihuman by dkappe with a higher blitz rating. Reading the description I think this can be easily adapted for CCRL. It has a limit of around 2500 nodes.
Mean Girl 8 (32x4) -- the most fun leela-style network -- looking at ~2500 nodes on a Raspberry Pi 3. Will play casual and rated Blitz, Rapid and Classical with at least 3 sec increment in both standard and chess960. Will move almost instantly. Designed to be a reasonable sparring partner. Uses a gambit book for extra entertainment value.
List of bots from lichess-bots team with 500 or more blitz games.
Thanks. Do any of these already have known approximate CCRL blitz ratings? If not I can probably manage to get the maia bots running on Lc0 at one node and test them myself.
Lichess bots are highly underrated. Thanks to all the SF bots that are running on the site and hammer the self-made bots.
But looking at Zahak for example even when it was around CCRL 2100 and running on a raspberry pi 3 (32 bit) it still could beat all human players that it has played.
