BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
I think the simplest solution in picking the strongest move against a human in terms of winning the game if you had a 32-man tablebase. Would be to play the move that keeps the game going the longest when it see a draw, then pick the shortest path to victory when it sees a win in the tablebase.
This longest-shortest strategy would be very hard to beat. And would drive a human crazy, causing a losing mistake in almost every game.
If you really want to be mean, have the tablebase play the longest path to victory in won positions.
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
I think the simplest solution in picking the strongest move against a human in terms of winning the game if you had a 32-man tablebase. Would be to play the move that keeps the game going the longest when it see a draw, then pick the shortest path to victory when it sees a win in the tablebase.
This longest-shortest strategy would be very hard to beat. And would drive a human crazy, causing a losing mistake in almost every game.
If you really want to be mean, have the tablebase play the longest path to victory in won positions.
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
Random moves are much stronger than losing move because a random player will draw once in a while even against a perfect player.
Hi Don,
the above statement doesn't make sense, as random moves are more likely to be bad, leading to a loss sooner rather than later, even if some random moves may actually be good along the way.
Interesting idea about calibrating to zero based on random moves, though.
Regards,
CL
I don't understand your difficulty. A random move might be the best move on the board and thus 2 random moves in a row have a chance to be two best moves in a row. 3 is even less likely but still possible and so on. Therefore it is theoretically possible for a random player to play a perfect game, i.e. every move is best. I never claimed it was likely, but the context of the argument was about playing infinitely bad and that is almost impossible to do unless you simply resign on every move.
Chances are that a random player will play horribly but due to the laws of probability it will occasionally play good enough to draw even a perfect player.
Of course I understand how rare this is, but it's not impossible.
Exactly. In fact, we can make some extreme estimations to show it.
D = S * ln(p) or
D = 2.3 * S * log(p) [Equation 1]
If one of the players play perfect, the probability to draw for the other is to play good enough moves every single time. If an average game lasts N moves and the average probability for "good enough" is "g", then
So, if we have to make a perfect move out of 40 (g = 1/40) for N=100 moves,
we get that the maximum ELO compared to a random player is ~66,000 and that is very extreme.
A less extreme number could be g = 1/20 and N = 60 which would give a maximum ELO of ~32,000
Bigger number than those ballpark number obtained by "back in the envelope calculations" are impossible.
Miguel
Most people have a backwards model of how chess strength works. It has nothing to do with you, it's all about your opponent. You cannot "go after" the half point, you have to wait for your opponent to give it to you while not making a blunder yourself.
There is really no such thing as a "great" move. You are never in a losing position and muster so much brain power that you create a win. You hear language like that in chess books that romanticize chess sometimes, especially the older book that fawn over the old masters. But that is not how it works.
So for a random player to play a "good" game we really mean that it is "lucky" enough to avoid all the game theoretical half or full point losses.
I think a lot of positions have multiple moves that are (theoretically) the same so it's not quite so hard as having to find 1 move out of 40 for 50 or 60 moves in row. In some endings you are shuffling pieces and there are few bad moves - most of the moves are adequate. But in many position there is only 1 move that must be played to avoid throwing away the win or draw.
One theoretical truth here is that if you are losing, you cannot play a bad move - unless you define ALL moves as bad of course.
So the rule is that you don't play good chess even though we often say that. What we really mean is that you play less bad moves that other guy.
Yes, from the simplified logistic Delta = 400*log(g^N) = 400*N*log(g). For average game N=60, g=20 and Delta ~ 30,000 Elo points, the distance from the perfect player to a random mover. And another similar quantity from a random mover to a perfect loser.
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
where I can download the 32 men database?
How much storage space do you have?
Capital punishment would be more effective as a preventive measure if it were administered prior to the crime.
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
where I can download the 32 men database?
How much storage space do you have?
Well, if i delete the unnecessary quotes from the talkchess posts from my firefox cache,
then I will get HUGE space to at least start with 32-men-TB download.
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
No reason to feel quilty.
Chess players should play the move that gives them the best practical chance and not the best theoretical move.
If you see that you objectively has a losing position then the target is not to lose in more moves so I see no problem if you play a move that is not the best from computer point of view even if you know it.
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
where I can download the 32 men database?
How much storage space do you have?
Well, if i delete the unnecessary quotes from the talkchess posts from my firefox cache,
then I will get HUGE space to at least start with 32-men-TB download.
A good one
_No one can hit as hard as life.But it ain’t about how hard you can hit.It’s about how hard you can get hit and keep moving forward.How much you can take and keep moving forward….
BubbaTough wrote:It sounds like not all perfect programs (defined as programs that, in any given position, always select one of the moves that preserves the best possible forceable outcome) will have the same rating, as there are different mechanisms for choosing between theoretically tied moves that will do better or worse at encouraging imperfect play by an adversary depending on their type of imperfect play.
It all kind of reminds me of when people talk about strategies for maximizing cheapo chances in known tablebase positions.
-Sam
That's clearly the case, getting the best possible results relies on opponent modeling, which is based on playing the opponent and not the board.
Imagine having a 32 man database. It could conceivably go immediately into a draw by repetition believing that it can do no better than a draw anyway. That would be a correct assumption if you were playing another database, but not if you were playing a weak player. A 32 man database might immediately simplify making it much more likely its opponent could draw too.
If you had a 32 man database a good strategy for selecting your moves would be to attach a conventional program to it - and at the root level prune away all the moves that do not give the best game theoretic results. Such a program would play pretty much like current programs do except that they would never make any mistake that would cost them 1/2 a point.
It's also possible that the best strategy would be to make a losing move against a particular opponent. Imagine the game ending for a sure draw but being able to complicate things by playing a sub-optimal move that leads to a loss against best play. Since you are playing the odds with opponent modeling you would want to play the move that was most likely to give you a win given what you know about your opponent. That may not always be the best move in the absolute sense although I doubt that playing a losing move would be a good choice very often.
I saw a strong player do this once to avoid a draw - he played a move that he actually estimated would lose the game against a good player but he was playing a much weaker player and a draw was almost certain and he could not afford a draw. The move created a much more complicated position that required a lot of technique to win and provided plenty of opportunities to go wrong.
I did something like that once myself, but I was probably going to lose anyway regardless of what I did, so in my case it was desperation. Instead of losing slow and sure I found an unsound sacrifice against a much weaker opponent who was beating me. The sacrifice had an "obvious" response which didn't work so it was a true swindle. I didn't expect my opponent to even fall for it because it smelled right from the start but after spending a really long time looking at it he played the bogus move and said, "you are just gonna have to show me." I won that game but felt guilty about winning it - like I had somehow cheated him. He was a super nice guy and after the game we laughed about it, but he made me admit that I was busted up until this point.
But that is opponent modeling. I have to wonder if that makes a difference in the strength of chess programs too. Chess program do not worry about their opponents (except in the case of contempt factors) but some are clearly more aggressive and other are more careful. Like humans they each have different styles which may work better or not as well against specific opponents or styles.
where I can download the 32 men database?
How much storage space do you have?
Well, if i delete the unnecessary quotes from the talkchess posts from my firefox cache,
then I will get HUGE space to at least start with 32-men-TB download.
Kai, I'm very fascinated with this post and am wondering if you would be willing to run another experiment. Jonathan Lee had replied that "If you did 20 years of Computer Chess, try Mephisto Genius version 2 or version 3." Would you be willing to run 100 or maybe 300 games of Shredder 7.04 against one of those chess programs?
Though I do have to add the caveats that I'm not sure how hard those versions are to find and whether or not it's a fair comparison because that software might not take full advantage of the hardware that it's run on whereas Shredder 7.04 may. Of course a match between Houdini 3 and Genius version 2 or 3 could be interesting, but that could take on the order of 10,000 games to produce a draw!!
Well, if i delete the unnecessary quotes from the talkchess posts from my firefox cache, 
ty!