A new game-by-game rating system

[Marc Andre Beaudry]

Your rating system does not make much sense to me ! Not very acurate.

[Marc Andre Beaudry]

How many Fischer moves would you have found in a 1400 game. Not many I guess, that's why it's called game of the century, You would have Be6 and other moves doubt it.

[Marc Andre Beaudry]

Not to mention Na4 !!

[Marc Andre Beaudry]

You put Fischer performance at 1400, meaning that 1400 players would find the moves of Fischer, I seriously doubt. Your rating system is interesting but you shouldn't put too much faith in it.
In the 31 st Karpov Korchnoi match, it is a very difficult rook endgame, where the 2 best rook endgame player of that period (to me the best rook endgame players are Karpov and Rubinstein) haver a go at it and you rate their performance extremely low. It is just rubbish seeing the great beauty and audacity of Korchnoi play and the quite tenacious defence of Karpov. This is a great endgame not a 1700 performance.

[Marc Andre Beaudry]

You've got to find a way to improve on your rating system, right now, it is useless, that's just my opinion, but you take 2 great game and reduce them to low ratings, that is far from an accurate picture.

[Pargat Perrer]

How many Fischer moves would you have found in a 1400 game. Not many I guess, that's why it's called game of the century, You would have Be6 and other moves doubt it.

You put Fischer performance at 1400, meaning that 1400 players would find the moves of Fischer, I seriously doubt. Your rating system is interesting but you shouldn't put too much faith in it.
In the 31 st Karpov Korchnoi match, it is a very difficult rook endgame, where the 2 best rook endgame player of that period (to me the best rook endgame players are Karpov and Rubinstein) haver a go at it and you rate their performance extremely low. It is just rubbish seeing the great beauty and audacity of Korchnoi play and the quite tenacious defence of Karpov. This is a great endgame not a 1700 performance.

You've got to find a way to improve on your rating system, right now, it is useless, that's just my opinion, but you take 2 great game and reduce them to low ratings, that is far from an accurate picture.

Marc, when I post the results of my GPR calculations, I make sure to say that these ARE NOT ELO NUMBERS. They are not to be considered as ELO ratings.

It's like if you live in the USA and regularly drive on the highway at the speed limit of 70 mph, and then you come to Canada and see the sign "Maximum Speed 100" on the 401, do you think you can drive 100 mph? It's two very different number systems.

So that is a mistake you are making. But... you are correct that the GPR numbers do show that the Karpiv vs Korchnoi (game 32 1978) and the D. Byrne vs Fischer 1956 games are poorly played. That is an absolutely correct assertion, and it doesn't matter your opinion on certain human players and how good they were in rook endgames... Stockfish would make mincemeat of either of them in rook endgames! Why do you think the top human GMs are not playing Stockfish? They'd be torn to pieces. Stockfish and other top engines are not allowed to participate in human tournaments, in order to save embarrassment for humanity.

The GPR is not useless: it exposes games that are otherwise considered great as being full of blunders. And it also shows that some human games, such as the Fischer vs Smyslov game that started this thread and that is not very well known or famous, are in fact the REAL great games.

This is part of what GPR is all about: saying something new about individual games. Not about career strength of players, but about individual games.

Fischer's ...Be6 and ...Na4 in his game against D. Byrne should actually be very easy to find moves. When I started Stockfish analyzing the positions where these 2 moves were played by Fischer, Stockfish put both these moves at the top WITHIN ONE SECOND! If some other move had taken the top spot and after several minutes Stockfish suddenly found very deep in the search tree that ...Be6 was suddenly good or ...Na4 was suddenly good, that would be different. But those moves were good immediately and stayed at the top the entire time of analysis. If you have Stockfish, you can prove it for yourself.

If we ran a test for human chess players, where they were presented with these 2 positions and (not recognizing them) something like 90% of them couldn't find ...Be6 or ...Na4 then we would have a story, because it means humans cannot find even the most obviously good moves in certain types of positions.

Marc, I'm sorry if this new thing I"m doing goes against your opinions of some very prominent human chess players. In your mind, you can continue to think of them as geniuses and you can continue to think if them as fantastic rook endgame players or you can continue to think of D.Byrne vs. Fischer 1956 as the game of the century. But raw calculations are showing otherwise.

[Robert Villeneuve]

Hello Pargat,

I don't know your level of chess playing. But to say things like 'because it means humans cannot find even the most obviously good moves in certain types of positions' strikes me as utter nonsense.

Engines see millions of moves in seconds. Humans usually consider only a few candidate moves. Your use of 'obviously good moves' is extremely biased towards engines. Moves are hard to find BECAUSE humans find them hard to find. Period. The fact that Stockfish, rated 3500, finds the move in one second does not make it an 'obviously good move'. It just makes it obvious for a chess engine. Trying to belittle Fischer's achievement, for example, by saying that a machine finds ...Na4 extremely quickly means nothing to the chess amateurs.

There's nothing obvious about computers and brute calculation. That's not how humans think.

You claim that the GPR is all about saying something new about individual games. In what way is this different from the little percentage analysis I get from chess.com after a game I play?

Every player, these days, enters their games and gets an engine of their choice to analyze them.

Sincerely,


Robert

[Pargat Perrer]

Hello Pargat,

I don't know your level of chess playing. But to say things like 'because it means humans cannot find even the most obviously good moves in certain types of positions' strikes me as utter nonsense.

Engines see millions of moves in seconds. Humans usually consider only a few candidate moves. Your use of 'obviously good moves' is extremely biased towards engines. Moves are hard to find BECAUSE humans find them hard to find. Period. The fact that Stockfish, rated 3500, finds the move in one second does not make it an 'obviously good move'. It just makes it obvious for a chess engine. Trying to belittle Fischer's achievement, for example, by saying that a machine finds ...Na4 extremely quickly means nothing to the chess amateurs.

Hi Robert, thanks for responding. I think we are in agreement at least on one point, we are just saying it different ways.

You are saying that moves are hard to find BECAUSE humans find them hard to find. Great, that is exactly what GPR is showing. It is showing that humans, even the ones who are best at chess, play bad chess. Here, good and bad is judged first by ELO rating and now by GPR numbers, and thanks for including the Stockfish 3500+ ELO rating. That's only about 700 points above Carlsen, the best human.

Earlier in this thread, Brad Thomson mentioned that in chess only winning matters. If that is true, then only chess engines play good chess. We may not like their style, it may not be like Tal very often, but we can't deny: computer engines rule the world of chess when it comes to winning and losing.

So if GPR is going to show is anything, it should be showing us that humans play bad chess. And that is exactly what it is showing. The reasons WHY don't matter, it doesn't matter that engines see millions of moves in one second. It only matters the ELO ratings.

By the way, you said I try to "belittle" Fischer's achievement. That is totally wrong. Belittle means to make fun of. That's an inflammatory accusation and I don't agree with it at all. I am being very objective in all of this. If Fischer is the only human who could have played ...Na4 or ...Be6 in those positions, then if there's any belittling to be done, it should be for the rest of the human players who wouldn't play those moves (but again, I am not belittling anyone). And that was the context where my quote was taken from, so there is no making fun of Fischer's achievement. I only say that the engines put both moves at the very top within 1 second. I'm stating an objective fact.

And don't forget, if a move is better according to the engines, that means it's better for winning. That's why chess moves get picked at all: to win.

There's nothing obvious about computers and brute calculation. That's not how humans think.

You claim that the GPR is all about saying something new about individual games. In what way is this different from the little percentage analysis I get from chess.com after a game I play?

First of all, we would need to have hundreds or even thousands of GPR rated games before we can compare individual games with each other. Just from the few I've done, we are seeing that Fischer vs. Smyslov was played TREMENDOUSLY better by both players than any other game I've done so far. But perhaps there are other games that will be found to be even much better played than that one. I expect some of the Fischer-Spassky games from 1972 to be very high, maybe some of the Kasparov-Anand games, and probably a whole lot of Carlsen's games because Carlsen is the one WC most influenced by computer engines.

Second, I don't know anything about what chess.com is doing, so maybe what I'm doing is not so original. Apparently Dr. Ken Regan did something like this back in 2012, but there's nothing new about it since then, and most of his work was in translating something like a GPR into an ELO number and making that transformation as accurate as possible. I don't see any necessity for that, GPR will stand on its own as a different number from ELO.

[Marc Andre Beaudry]

Many times a +10 move is more complicated than a mere +3 that the human would choose. Computers are just simply way way above us in calculation. By the way, we know since 1997 that computers are much better than us :-))

[Pargat Perrer]

Many times a +10 move is more complicated than a mere +3 that the human would choose. Computers are just simply way way above us in calculation. By the way, we know since 1997 that computers are much better than us :-))

True, but my point with GPR is not to shout from the rooftops: "Computers are better than humans at chess!" I am only emphasizing it here in response to the idea that, for example, Karpov is so good at rook endgames that he can't possibly have made lots of mistakes in rook endgames. Well.... yes, he could have and he did.

The only reason something like a GPR is even possible is the fact that computers are so much better than us. To score the moves of a game, we need an ultimate authority. We can't ask God, so we do the next best thing: we use a "virtual" ultimate authority which is the computer engine. Perhaps in 25 years there will be an Alpha Zero for smartphones that will be playing 5000 ELO chess and make Stockfish 13 look silly. I actually doubt it will be that extreme though, because there are a very limited number of reasonable move choices in any position.

Some of my chess variants have a LOT more move choices. It would be great to be able to modify Stockfish 13 to play some variants, but I don't have the knowledge needed to do that. I have modified a 2000 ELO level chess engine to play some variants, but it is run by scripts, and I can't modify the innards of the engine itself.

[Pargat Perrer]

With one or two exceptions, I've been doing the GPR rating of mostly past-generation players who excelled at chess before the advent of computers. Now here's 2 players that came to prominence during the rise of computer engines: Kasparov and Anand.

This is game 10 of their WC match in Manhattan in 1995, just 2 years before Kasparov was to lose to Deep Blue.

The first 8 games of the match were all draws. Then Anand broke through for a decisive Game 9 victory. I could have rated that game, but since Kasparov ended up winning the title, I thought he deserved to have one of his victories done instead.

Here it is and it's a whopper. I mean, Kasparov takes a flying leap past Fischer's (so-far) highest GPR ever. GPR rating is still in its infancy, and I have a few ideas of how to modify it a bit, so having the highest GPR doesn't mean so much yet, but I have a feeling this one will stay at the top for a while.

Here's the pgn...

[Date "1995-09-26"]
[Result "1-0"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"]
[White "Garry Kasparov"]
[Black "Viswanathan Anand"]
[Event "PCA World Championship Match (1995)"]
[Site "New York, NY USA"]
[Round "10"]
1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Nxe4 6.d4 b5 7.Bb3 d5 8.dxe5 Be6 9.Nbd2 Nc5 10.c3 d4 11.Ng5 dxc3 12.Nxe6 fxe6 13.bxc3 Qd3 14.Bc2 Qxc3 15.Nb3 Nxb3 16.Bxb3 Nd4 17.Qg4 Qxa1 18.Bxe6 Rd8 19.Bh6 Qc3 20.Bxg7 Qd3 21.Bxh8 Qg6 22.Bf6 Be7 23.Bxe7 Qxg4 24.Bxg4 Kxe7 25.Rc1 c6 26.f4 a5 27.Kf2 a4 28.Ke3 b4 29.Bd1 a3 30.g4 Rd5 31.Rc4 c5 32.Ke4 Rd8 33.Rxc5 Ne6 34.Rd5 Rc8 35.f5 Rc4+ 36.Ke3 Nc5 37.g5 Rc1 38.Rd6 1-0

And here were the GPR results:

Kasparov Game Performance Rating (GPR): 10127
Anand Game Performance Rating (GPR): 1897

That is total domination, and Kasparov went on to win 2 of the next 3 games as Anand's confidence must have taken a huge hit from this game, perhaps his worst game ever at the top levels. This game was perhaps the most critical of Kasparov's entire career, his reputation truly on the line.

[Sid Belzberg]

With one or two exceptions, I've been doing the GPR rating of mostly past-generation players who excelled at chess before the advent of computers. Now here's 2 players that came to prominence during the rise of computer engines: Kasparov and Anand.

This is game 10 of their WC match in Manhattan in 1995, just 2 years before Kasparov was to lose to Deep Blue.

The first 8 games of the match were all draws. Then Anand broke through for a decisive Game 9 victory. I could have rated that game, but since Kasparov ended up winning the title, I thought he deserved to have one of his victories done instead.

Here it is and it's a whopper. I mean, Kasparov takes a flying leap past Fischer's (so-far) highest GPR ever. GPR rating is still in its infancy, and I have a few ideas of how to modify it a bit, so having the highest GPR doesn't mean so much yet, but I have a feeling this one will stay at the top for a while.

Here's the pgn...

[Date "1995-09-26"]
[Result "1-0"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"]
[White "Garry Kasparov"]
[Black "Viswanathan Anand"]
[Event "PCA World Championship Match (1995)"]
[Site "New York, NY USA"]
[Round "10"]
1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Nxe4 6.d4 b5 7.Bb3 d5 8.dxe5 Be6 9.Nbd2 Nc5 10.c3 d4 11.Ng5 dxc3 12.Nxe6 fxe6 13.bxc3 Qd3 14.Bc2 Qxc3 15.Nb3 Nxb3 16.Bxb3 Nd4 17.Qg4 Qxa1 18.Bxe6 Rd8 19.Bh6 Qc3 20.Bxg7 Qd3 21.Bxh8 Qg6 22.Bf6 Be7 23.Bxe7 Qxg4 24.Bxg4 Kxe7 25.Rc1 c6 26.f4 a5 27.Kf2 a4 28.Ke3 b4 29.Bd1 a3 30.g4 Rd5 31.Rc4 c5 32.Ke4 Rd8 33.Rxc5 Ne6 34.Rd5 Rc8 35.f5 Rc4+ 36.Ke3 Nc5 37.g5 Rc1 38.Rd6 1-0

And here were the GPR results:

Kasparov Game Performance Rating (GPR): 10127
Anand Game Performance Rating (GPR): 1897

That is total domination, and Kasparov went on to win 2 of the next 3 games as Anand's confidence must have taken a huge hit from this game, perhaps his worst game ever at the top levels. This game was perhaps the most critical of Kasparov's entire career, his reputation truly on the line.

Pargat I think the GPR idea is an awesome one. Thanks for sharing.

[Pargat Perrer]

Pargat I think the GPR idea is an awesome one. Thanks for sharing.

Thank you, Sid, I am glad that someone is appreciating what these numbers are telling us about past games.

In order to avoid the confusion with ELO numbers (for example, a GPR of 1800 being confused with an ELO rating of 1800), I am instuting a minor change: I'm going to divide the calculated GPR by 10, rounding up the resulting decimal digit so that .50 or above adds a digit, .49 or below doesn't, and that becomes the new GPR number.

So here's a summary of the GPRs I have calculated so far, with this new divide-by-10 implementation:

Fischer vs Smyslov 1965:
Fischer Game Performance Rating (GPR): 857
Smyslov Game Performance Rating (GPR): 427


Hou Yifan vs. Judit Polgar 2012:
Yifan Game Performance Rating (GPR): 316
Polgar Game Performance Rating (GPR): 284


Botvinnik vs. Tal 1960 Game 6:
Botvinnik Game Performance Rating (GPR): 146
Tal Game Performance Rating (GPR): 180


Zhe Quan vs. Jean Hebert 2005:
Quan Game Performance Rating (GPR): 29
Hebert Game Performance Rating (GPR): 303


Adolf Anderssen vs. Wilhelm Steinitz 1866:
Anderssen Game Performance Rating (GPR): 613
Steinitz Game Performance Rating (GPR): 186


Korchnoi vs. Karpov 1978 Game 31:
Korchnoi Game Performance Rating (GPR): 542
Karpov Game Performance Rating (GPR): 276


Karpov vs. Korchnoi 1978 Game 32:
Karpov Game Performance Rating (GPR): 210
Korchnoi Game Performance Rating (GPR): 135


D.Byrne vs. R.Fischer 1956:
Byrne Game Performance Rating (GPR): 71
Fischer Game Performance Rating (GPR): 143


Kasparov vs. Anand 1995 Game 10:
Kasparov Game Performance Rating (GPR): 1013
Anand Game Performance Rating (GPR): 190


and the lone Rapid Chess game....

Carlsen vs Giri Rapid Chess 2021:
Carlsen Game Performance Rating - Rapid (GPR-R): 124
Giri Game Performance Rating - Rapid (GPR-R): 106

[Pargat Perrer]

In doing my next set of GPR ratings (more on that later), I came across one of those annoying games in which neither player shows any intention of playing for a win. Usually these games become agreed draws by move 20 or so. Perhaps in this event there was a rule against that, because the players didn't agree to a draw until Black's move 57. But there's no denying that neither player ever tried for a win; perhaps each was just hoping the other would make a key mistake.

This is problematic for GPR ratings, for one simple reason: chess engines rate drawish moves very highly, including tradeoff of equal value pieces in which the player making the RECAPTURE has multiple ways of doing it, one or more of which may be a mistake.

It is reasonable that engines would give such moves high values. They are not mistakes. And therein lies the root problem with competitive chess: it relies on mistakes being made. Without mistakes, chess is nothing but one boring draw after another. There are no brilliant moves that force a win in an equal position. There are only moves that INDUCE MISTAKES, and making such a move could be considered brilliant simply for the fact that it induced a mistake. If it didn't induce a mistake, then it better not be a mistake on its own.

So in the case of a player like Tal, as Brad Thomson mentioned, he played a lot of moves trying to induce mistakes, but his record shows that a lot of the moves he made were mistakes in their own right and when they didn't induce a corresponding and larger mistake, they cost Tal games. But we admire Tal because he tried courageously to win, and that is the essence of competitive chess.


The only solution for GPR ratings is to introduce some subjectivity by allowing the entity doing the GPR rating (at this time, that's me) to DECLARE that neither player is playing for a win, and therefore the ENTIRE GAME is ineligible for GPR ratings. In the case I've just mentioned, if I did allow the game to be rated, both player would have a GPR in excess of 10,000 which is absurd.

Making a game ineligible for GPR rating does carry an inherent penalty. If 2 or more players have an equal GPR, the official ranking of those players is based on the most GPR rated moves. So if you play an intentional draw and lose all those moves to GPR rating, you will suffer in any tie-breaking.


Now, in the interest of full disclosure and since I am the entity that is being subjective in declaring a game is ineligible for GPR....

I am doing GPR rating for the entire event that just finished, the Sinquefield Cup. I am still doing games from the first round, and there are 9 rounds, so it will be several weeks before I post the results of the ratings.

But the game in question in the first round is the 57-move draw between Xiong vs. Rapport. Here is the pgn:

[Date "2021-08-17"]
[Result "½-½"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"]
[White "Jeffery Xiong"]
[Black "Richard Rapport"]
[Event "Sinquefield Cup 2021"]
[Site "St. Louis, MO USA"]
[Round "1"]
1.d4 Nf6 2.c4 e6 3.Nc3 Bb4 4.Qc2 c5 5.dxc5 Qa5 6.Bd2 Qxc5 7.e3 d5 8.a3 Bxc3 9.Bxc3 dxc4 10.Bb4 Qc6 11.Qxc4 Qxc4 12.Bxc4 Nc6 13.Bc3 Ne4 14.Rc1 Nxc3 15.Rxc3 Bd7 16.Nf3 Rc8 17.Bd3 Ke7 18.Ke2 h6 19.Rhc1 Nb8 20.Ne5 Rxc3 21.Rxc3 Rc8 22.Kd2 Rxc3 23.Kxc3 Bc8 24.f4 Nd7 25.Nxd7 Bxd7 26.e4 f6 27.Kd4 Kd6 28.e5+ fxe5+ 29.fxe5+ Ke7 30.Be4 b6 31.Kc4 Kd8 32.Kd4 Ke7 33.h4 Bb5 34.g3 Kd7 35.Bc2 Be2 36.Ba4+ Kc7 37.Ke3 Bh5 38.Kf4 Be2 39.g4 Kd8 40.Bc6 Ke7 41.g5 Bh5 42.b4 Bd1 43.a4 Bh5 44.Bb5 Bd1 45.a5 Bh5 46.Bf1 Be8 47.Bg2 Bh5 48.Bc6 Bd1 49.gxh6 gxh6 50.Ke3 Bh5 51.Kd3 Kd8 52.axb6 axb6 53.Ke3 Ke7 54.Kd3 Kd8 55.Ke3 Ke7 56.Kd3 Kd8 57.Ke3 ½-½

If you think I am wrong in declaring this game to be too drawish to be eligible for GPR ratings, please post and give your reasons.

[Brad Thomson]

It seems to me that your analysis is in the final analysis aesthetic, which is not a bad thing. Tal's provocative moves, while very aesthetically pleasing and perhaps therefore worthy of a high score, are not as far as a computer is concerned as good as drawish exchanges. So we have a question of artistic taste rather than pure logic, and again this is not bad. The issue is that we all have differing aesthetic sensibilities. I prefer Duane Allman, others may prefer Jimi Hendrix (though I love Hendrix), I prefer Rachmaninoff, others may prefer Tchaikovsky, I prefer Bobby Orr, others may prefer Wayne Gretzky...