5 ms·

An interesting tidbit is that while the difference in rating between Carlsen and Anand is 95 points, the difference between Carlsen and Houdini 3 is 376 points.

by jules 13y ago

gautamnarula 13y ago
Another interesting tidbit is that the difference in rating between the top ranked player (Carlsen) and the World Champion (Anand) is the largest since Spassky-Fischer in 1972, when Champion Spassky was rated 2660 and Fischer was rated 2785--a 125 point difference!
- asdfologist 13y ago
  At least some (if not all) of that can be explained by rating inflation. The same gap in ability translates to many more points today than 40 years ago.
  - billforsternz 13y ago
    >The same gap in ability translates to many more points today than 40 years ago Not really. Ratings inflation means you can't compare the absolute values, but is much less noticeable on ratings differences. So 2660 + 4.7% = 2785 (125 points) in 1972. The equivalent today would be something like; 2750 + 4.7% = 2879 (129 points) The same gap in ability translates to only 4 more points. Actually even that modest difference should not arise from ratings inflation. That is because the Elo system is based on the idea that a given "gap in ability" is supposed to yield a constant Elo delta. An Elo delta of 100 points is supposed to translate to a 64% score for the stronger player, a delta of 200 points to 76% etc.
    - asdfologist 13y ago
      Interesting, that sounds reasonable, but two questions: 1. Is it certain that rating inflation is linear? 2. Is there actually any evidence that the constant Elo delta property actually holds even with inflation? It's well known that Elo is flawed, which is why people have been working on alternate rating systems...
      - billforsternz 13y ago
        I am not a ratings expert, just a (serious) player, chess programmer and amateur chess journalist. I am not sure ratings inflation even exists. If it does exist and you are talking about rating differences between world class players, then you are talking about small differences in big numbers and you are going to get a good approximation to linearity. But it's only approximate because in my previous example the linearity yielded 125 points v 129 points when the key underlying property of the system requires 125 points v 125 points. But unfortunately I don't know the answer to your question 2), I don't know how robust that property of the system is. Why do I say ratings inflation may not exist ? It is true that in Fischer's time there was one 2700+ player. Now there are 30 or so. That's the evidence for ratings inflation. But it seems likely that the top players are just better than they used to me. Maybe there really are 30 players today who can play at 1972 Bobby Fischer level.
  - gautamnarula 13y ago
    One statistician has developed a system to compare players across time. I don't know enough about statistics to say if it's valid or not, but it's interesting to look at. Unfortunately, it hasn't been updated in almost a decade, since it says Carlsen's peak rating was 2556 in January 2005. http://www.chessmetrics.com/cm/ http://www.chessmetrics.com/cm/
- iends 13y ago
  Anand is older and has less incentive to play in open tournaments when he gets a huge payout for defending his title every 1-2 years. He's playing just enough chess to stay sharp, and saving all his analysis/novelties/hard work for the big payout of the world championship. He simply doesn't care about rating.
zwegner 13y ago
Elo scores in different rating pools are not directly comparable at all, only the difference in ratings in the same pool is significant (although generally computer chess rating lists try to make them look reasonably close to the FIDE list). More big problems in comparing these ratings: intransitivity of chess strength, variation of strength for different time controls, volatility/time variation of human strength, and the much lower statistical significance of human rating lists. There's probably more that don't come to mind. But anyways, yeah, it's pretty indisputable that computers absolutely massacre humans at chess these days.
adamconroy 13y ago
The problem with this assessment is that computers basically cheat. Carlsen isn't allowed to turn up to a game with an openings book, and some way of looking up all possible 7 piece or less endgames.(if I remember correctly), but the computer can. Another way of looking at is that if Carlsen had all that information at his disposal every time he played another human, his rating would be 300+ higher than everyone else as well.
- shaggyfrog 13y ago
  This is not cheating in any real sense of that word. The computer has that opening book in its memory, so why shouldn't it make use of it any more than Carlsen is making use of his memory? What's the alternative? Harrison Bergeron-style handicapping of computers, to soothe our human egos?
- tzs 13y ago
  Today's computers would still dominate even if their opening books were taken away. Open book essentially represents a very extensive analysis of one position (the starting position) that has been carried out by humans over centuries. We aren't as fast as computers, but we've been working on this one problem longer and with massive parallelism, and so we've got the main lines covered quite deep. It used to be that most of book went deeper than a bookless computer could analyze in the time available for the opening in a timed game. This meant that a booked up human could steer the game into lines that appeared favorable for the computer when analyzed to the depth the computer could analyze, but were actually favorable to the human when analyzed to book depth. Computers are now fast enough that they can reach or surpass book depth in most lines, making it much much harder for a booked up human to steer the game into a favorable line. As an example of how fast computers are now, I just let Stockfish 4 64bit have a go at the start position on my 2008 Mac Pro (dual 2.8 GHz quad core Xeon processors, without hyper threading). I gave it 3 minutes (3 minutes per move is common in world class chess events). These were the settings: 8 threads, idle threads do not sleep, 8 GB hash table. It reached a depth of 30 ply (44 ply in selective search), and examined 940 million nodes, averaging around 5.3 million nodes per second. (It picked d4 as its move, and expects a reply of d5). The endgame tables also aren't as important as you might think at first. They only cover endgames without pawns, and so it is not that often that positions from these tables appear in a game. If one of these positions DOES show up in a game, and the computer misplays it without access to the tables, that isn't likely to give the human the win--it most likely will just let the human draw in a lost position. That's because the problem computers have with these endings without using tables is that they aren't able to see how to make progress from a won position so they end up hitting the 50 move draw rule. (Note that this only applies to games with a lot of pieces left. Even fairly old chess programs could do the basic piece-only endgames without needing tables). Thus, if the tables were taken away, the overall result would be that in the games where pawns remain (most games) computers would win a lot more than humans, and in games where all the pawns go away while there are still several pieces left humans would get some draws in lost positions. Overall, take away both opening book and endgame tables, and computers would still dominate, just not by quite as much because of the rare pieces only endgame that humans would sometimes draw from a lost position, and the rare case where the human could steer the opening into one of the few lines where book is still deeper than the computer can reach and that line's evaluation changes between the depth the computer reaches and the depth the book reaches.

An interesting tidbit is that while the difference in rating between Carlsen and Anand is 95 points, the difference between Carlsen and Houdini 3 is 376 points.

by jules 13y ago

An interesting tidbit is that while the difference in rating between Carlsen and Anand is 95 points, the difference between Carlsen and Houdini 3 is 376 points.

gautamnarula 13y ago

Another interesting tidbit is that the difference in rating between the top ranked player (Carlsen) and the World Champion (Anand) is the largest since Spassky-Fischer in 1972, when Champion Spassky was rated 2660 and Fischer was rated 2785--a 125 point difference!

asdfologist 13y ago
At least some (if not all) of that can be explained by rating inflation. The same gap in ability translates to many more points today than 40 years ago.
- billforsternz 13y ago
  >The same gap in ability translates to many more points today than 40 years ago Not really. Ratings inflation means you can't compare the absolute values, but is much less noticeable on ratings differences. So 2660 + 4.7% = 2785 (125 points) in 1972. The equivalent today would be something like; 2750 + 4.7% = 2879 (129 points) The same gap in ability translates to only 4 more points. Actually even that modest difference should not arise from ratings inflation. That is because the Elo system is based on the idea that a given "gap in ability" is supposed to yield a constant Elo delta. An Elo delta of 100 points is supposed to translate to a 64% score for the stronger player, a delta of 200 points to 76% etc.
  - asdfologist 13y ago
    Interesting, that sounds reasonable, but two questions: 1. Is it certain that rating inflation is linear? 2. Is there actually any evidence that the constant Elo delta property actually holds even with inflation? It's well known that Elo is flawed, which is why people have been working on alternate rating systems...
    - billforsternz 13y ago
      I am not a ratings expert, just a (serious) player, chess programmer and amateur chess journalist. I am not sure ratings inflation even exists. If it does exist and you are talking about rating differences between world class players, then you are talking about small differences in big numbers and you are going to get a good approximation to linearity. But it's only approximate because in my previous example the linearity yielded 125 points v 129 points when the key underlying property of the system requires 125 points v 125 points. But unfortunately I don't know the answer to your question 2), I don't know how robust that property of the system is. Why do I say ratings inflation may not exist ? It is true that in Fischer's time there was one 2700+ player. Now there are 30 or so. That's the evidence for ratings inflation. But it seems likely that the top players are just better than they used to me. Maybe there really are 30 players today who can play at 1972 Bobby Fischer level.
- gautamnarula 13y ago
  One statistician has developed a system to compare players across time. I don't know enough about statistics to say if it's valid or not, but it's interesting to look at. Unfortunately, it hasn't been updated in almost a decade, since it says Carlsen's peak rating was 2556 in January 2005. http://www.chessmetrics.com/cm/ http://www.chessmetrics.com/cm/
iends 13y ago
Anand is older and has less incentive to play in open tournaments when he gets a huge payout for defending his title every 1-2 years. He's playing just enough chess to stay sharp, and saving all his analysis/novelties/hard work for the big payout of the world championship. He simply doesn't care about rating.

zwegner 13y ago

Elo scores in different rating pools are not directly comparable at all, only the difference in ratings in the same pool is significant (although generally computer chess rating lists try to make them look reasonably close to the FIDE list). More big problems in comparing these ratings: intransitivity of chess strength, variation of strength for different time controls, volatility/time variation of human strength, and the much lower statistical significance of human rating lists. There's probably more that don't come to mind. But anyways, yeah, it's pretty indisputable that computers absolutely massacre humans at chess these days.

adamconroy 13y ago

The problem with this assessment is that computers basically cheat. Carlsen isn't allowed to turn up to a game with an openings book, and some way of looking up all possible 7 piece or less endgames.(if I remember correctly), but the computer can. Another way of looking at is that if Carlsen had all that information at his disposal every time he played another human, his rating would be 300+ higher than everyone else as well.

shaggyfrog 13y ago
This is not cheating in any real sense of that word. The computer has that opening book in its memory, so why shouldn't it make use of it any more than Carlsen is making use of his memory? What's the alternative? Harrison Bergeron-style handicapping of computers, to soothe our human egos?
tzs 13y ago
Today's computers would still dominate even if their opening books were taken away. Open book essentially represents a very extensive analysis of one position (the starting position) that has been carried out by humans over centuries. We aren't as fast as computers, but we've been working on this one problem longer and with massive parallelism, and so we've got the main lines covered quite deep. It used to be that most of book went deeper than a bookless computer could analyze in the time available for the opening in a timed game. This meant that a booked up human could steer the game into lines that appeared favorable for the computer when analyzed to the depth the computer could analyze, but were actually favorable to the human when analyzed to book depth. Computers are now fast enough that they can reach or surpass book depth in most lines, making it much much harder for a booked up human to steer the game into a favorable line. As an example of how fast computers are now, I just let Stockfish 4 64bit have a go at the start position on my 2008 Mac Pro (dual 2.8 GHz quad core Xeon processors, without hyper threading). I gave it 3 minutes (3 minutes per move is common in world class chess events). These were the settings: 8 threads, idle threads do not sleep, 8 GB hash table. It reached a depth of 30 ply (44 ply in selective search), and examined 940 million nodes, averaging around 5.3 million nodes per second. (It picked d4 as its move, and expects a reply of d5). The endgame tables also aren't as important as you might think at first. They only cover endgames without pawns, and so it is not that often that positions from these tables appear in a game. If one of these positions DOES show up in a game, and the computer misplays it without access to the tables, that isn't likely to give the human the win--it most likely will just let the human draw in a lost position. That's because the problem computers have with these endings without using tables is that they aren't able to see how to make progress from a won position so they end up hitting the 50 move draw rule. (Note that this only applies to games with a lot of pieces left. Even fairly old chess programs could do the basic piece-only endgames without needing tables). Thus, if the tables were taken away, the overall result would be that in the games where pawns remain (most games) computers would win a lot more than humans, and in games where all the pawns go away while there are still several pieces left humans would get some draws in lost positions. Overall, take away both opening book and endgame tables, and computers would still dominate, just not by quite as much because of the rare pieces only endgame that humans would sometimes draw from a lost position, and the rare case where the human could steer the opening into one of the few lines where book is still deeper than the computer can reach and that line's evaluation changes between the depth the computer reaches and the depth the book reaches.