# Are engine ratings directly comparable to human-level play?

Yesterday, I play a game against Stockfish 7, and it has 2200 rating points.

I drew 2 games, and lost 1. The point is that if I play against a human-rated 2200 player, I would probably have a positive score against him.

I also noticed that Stockfish 7 played much better than the average 2200 human player strength.

Do human-ratings and machine-ratings show the same strength of play?

• Hi, this post seems to have received a number of decent answers, if you have found one to be particularly satisfactory please consider accepting it, as it's important to give closure to well addressed posts. Thanks for considering it. Sep 12, 2019 at 11:50

What you have to understand is, first, that any rating is just a comparison of players within that pool. So a 2200 on chess.com is not the same as 2200 on lichess.com, or on ICC. They are probably close, but it is not the same.

Second, you need to realize that your three game sample is just not enough to draw any valid conclusions. You are human, so even if your winning expectancy were .95, you can always defy expectations, and make a blunder, and the computer, even handicapped, would still likely take advantage of it.

Lastly, at 2200, you are playing a handicapped computer. It is playing "not best" moves, but how is it playing them? Is it being told to play the fourth best move every time, or is it being told to play the best moves every time, but on the 5th move, play the 10th best move? Either way, you could have an anomaly where it plays less bad moves in an individual game, and when those worse moves do pop up, maybe you miss the opportunity.

The computer can only be handicapped so much.

• The computer didnt seem to blunder in any of the games. Sep 6, 2019 at 15:29
• Update: I played 9 more games and the score is for me: +1 , =10 , -1 Sep 6, 2019 at 23:40
• I cant play like a 2200 player due to the fact my rating is 2300. Sep 8, 2019 at 14:07
• @SoftwarePlayer: that doesn't follow, my rating is 1900-something and sometimes I can play like a 1200. Sep 9, 2019 at 10:02
• @RemcoGerlich Thanks. I missed that. I am skeptical based on his strange comments. Sep 9, 2019 at 10:13

There are several things to note:

• Computer ratings are inaccurate because there are so few serious games played between humans and computers that can be used to tune the ratings. Then the rest of the computer ratings are derived from computer v computer games. So it's 2200 plus or minus an error margin that could be hundreds of points.

• I have no idea where you got the rating number from, but I suspect it's a rounded number and may even be a pure guess by the engine author

• You have a FIDE rating of 2300, achieved in FIDE rated over the board tournaments. But do you play at exactly the same level against the computer? For me at least, my concentration level in a tournament hall playing for real results is way higher than at home in the evening clicking some pieces against a computer. The computer doesn't care, it is always the same.

• Ratings hide the fact that style differences exist. Your expected score against a 2200 player is 64%; but there are probably players of that level whose style is hard for you to counter and you do less well against them, there are also players against which you'll score better. The 64% is just an average.

• Especially the fact that you have 10 draws in 12 games suggests the combination of your style and the computer's leads to drawishness. And for instance part of the computer's 2200 rating may be in the fact that while it can draw 2300s a lot, it would also draw against 2100s more than you'd expect.

• Your hardware may be different than the hardware used to get the rating of 2200

• And in the end, a 50% score over 12 games is not enough games to conclude that the rating difference is not 100 (which would yield a 64% score over an infinity of games). There is always some variance in results and 12 games is not a lot.

So overall I think your 50% score is well within the range of normal results, as that range is broad.

If you are rated 2300, the computer is 2200 and you make 1/3, that's not too weird of a result, specially not if you played two of those three games as Black, for instance.

If you played 300 games and socred only 100 points, then you can start to have some suspicion that perhaps the engine is stronger than it says.

Even then, it could also be the case that you perform badly against that particular opponent (the engine), just as it could happen against any human player

This depends on how those engine ratings you're thinking of are generated in the first place. Ratings are reliable only so much as to compare vs. other players being rated. For example you say Stockfish 7 has rating 2200, but the CCRL 40/40 list gives a rating of 3326. Other problems: the strength of an engine depends on the hardware used, and to a lesser extent on the time control.

That said, TCEC attempted to benchmark its ratings to the FIDE ratings recently. The idea was to use Kramnik's rating when he played Fritz in Bahrain in 2002 to anchor Fritz in Bahrain's rating. That rating can then be used for every other engine running on the same hardware.

Based on that, Lc0's rating is currently 3907 (on TCEC hardware), and Stockfish dev is 3895.

• In lichess.org stockfish 7 level 7 has a rating of 2200. Sep 9, 2019 at 14:29

From the Stockfish FAQ, it is basically impossible to compare human and engine ELO, which is highly dependent on rating pool (which other engines played), hardware, time control, and choice of opening book (otherwise computer chess at the highest levels is 100% draws), as well as computers playing chess very differently than humans do. The only certainty is engines are much, much stronger than humans.

In 2016, Erik Varend tried a direct comparison of CCRL and FIDE rating systems. The conclusion is there is simply not enough data to draw any meaningful extrapolation. But for the sake of argument, Kai Laskos gives this rule of thumb:

FIDE rating = 0.70*CCRL rating + 840 ELO points

So Stockfish 16 at 3600 CCRL could be roughly 3360 FIDE (but it could be hundreds of points off).

Ratings depend on the time limit, the players, the method of rating. They are not comparable unless the factors are all the same. They do not cross boundaries meaningfully. And all of them are just statistical approximations of how good a player is at a given point in time. Players improve, some grow older and and do not play as well.