Time control is irrelevant to the perfect engine. If you can raise your Elo beyond an asymptotic limit by using ten times as much time, you've defeated your own argument about reaching the limit. The perfect engine must get a 50% score against itself using 100X more time. By this measure, current engines are far from perfect.
100x more time doesnt hold true across the entire time spectrum. the elo delta for an engine given 1s/move against itself given 100s/move, is gonna be much larger than if it were 1yr/move vs. 100yrs/move.
likewise, an optimal engine's elo differential against "the field" (e.g. top-5 strongest centaurs that exist) will be much smaller if the time controls were 1move/2days vs. 40/120.
likewise, an optimal engine's elo differential against "the field" (e.g. top-5 strongest centaurs that exist) will be much smaller if the time controls were 1move/2days vs. 40/120.
100x more time doesnt hold true across the entire time spectrum. the elo delta for an engine given 1s/move against itself given 100s/move, is gonna be much larger than if it were 1yr/move vs. 100yrs/move.
This is true, but it only shows that current chess engine improvement in Elo increases more slowly than logarithmic time. This says nothing about being near the draw limit.
likewise, an optimal engine's elo differential against "the field" (e.g. top-5 strongest centaurs that exist) will be much smaller if the time controls were 1move/2days vs. 40/120.
The optimal engine would beat the top-5 strongest centaurs that exist every single time except when the centaurs played an unflawed game, which would only occur an infinitesimal percentage of the time. So the time control would be irrelevant. This would be just as true if you gave the centaurs a month for every move. They still wouldn't be able to see continuations that are hundreds of moves long. But you won't need a perfect player to convincingly beat the five best centaur teams today. The Rybka cluster will do just fine.
This is true, but it only shows that current chess engine improvement in Elo increases more slowly than logarithmic time. This says nothing about being near the draw limit.
likewise, an optimal engine's elo differential against "the field" (e.g. top-5 strongest centaurs that exist) will be much smaller if the time controls were 1move/2days vs. 40/120.
The optimal engine would beat the top-5 strongest centaurs that exist every single time except when the centaurs played an unflawed game, which would only occur an infinitesimal percentage of the time. So the time control would be irrelevant. This would be just as true if you gave the centaurs a month for every move. They still wouldn't be able to see continuations that are hundreds of moves long. But you won't need a perfect player to convincingly beat the five best centaur teams today. The Rybka cluster will do just fine.
i heard freestyle chess isnt dead. we'll see just how many wins the cluster racks up in the next finals.
![[us]](/mwf/v12/flags/us.png "United States")
It may not be dead, but it isn't breathing and it doesn't have a pulse.
Gold is only useful as something to be admired for its uselessness!
And, yet in its "uselessness" , it has had the symbolic force to shape cultures down through the ages. How useless is that!
And, yet in its "uselessness" , it has had the symbolic force to shape cultures down through the ages. How useless is that!
![[mx]](/mwf/v12/flags/mx.png "Mexico")
Impressive sid! I've always been annoyed of the very short space that the forum gives to signatures (100 chars), many times the quote I want to put just doesn't fit there, and you managed to put two quotes in there!
On the other hand, if you shorten the time control, a truly perfect player will get more wins against players who are not perfect, thus gaining more elo, while with longer time controls, players who are not perfect at short time controls might become perfect much of the time at long time controls.
If you shorten the time control, the perfect player will still be perfect, but the non-perfect player will get worse, so the Elo gap will get larger (if you could measure it).
In order to play perfectly, you will need to be able to see continuations that are hundreds or thousands of moves long (as shown by certain 6+ man TBs, where this is required to get the half point). I'm pretty sure that we are far, far away from being able to see continuations that are hundreds of moves long without TBs...
In order to play perfectly, you will need to be able to see continuations that are hundreds or thousands of moves long (as shown by certain 6+ man TBs, where this is required to get the half point). I'm pretty sure that we are far, far away from being able to see continuations that are hundreds of moves long without TBs...
We agree on the first point, though not as much on the second, as we both know. :-) I feel that there will come a certain point where a program with a good search and evaluation will be able to find the most optimal moves the vast majority of the time at tournament time controls. I think this will happen within 5 years if it hasn't already happened with the Cluster.
"I think this will happen within 5 years if it hasn't already happened with the Cluster."
a bit off topic, but what i'd like to see is how well the cluster navigates tactical minefields w/o a deep book. i recall a post by jeroen that said certain lines in the PP(?) would require r3 to look 50ply ahead to see what's going on (his implication was unless u're well booked, no engine can survive the line). i only wish lukas would test jeroen's hypothesis to see if it's really true. it'd be a good (tactical) test for the cluster, and maybe vas could learn something about it to make the search/eval even stronger as a result.
a bit off topic, but what i'd like to see is how well the cluster navigates tactical minefields w/o a deep book. i recall a post by jeroen that said certain lines in the PP(?) would require r3 to look 50ply ahead to see what's going on (his implication was unless u're well booked, no engine can survive the line). i only wish lukas would test jeroen's hypothesis to see if it's really true. it'd be a good (tactical) test for the cluster, and maybe vas could learn something about it to make the search/eval even stronger as a result.
Strong as it is, I'm not even sure the cluster could find the short sequence that Eros used to beat Wayne. I see these kind of issues all the time in my cc games and don't expect them to go away in the next 5 years.
No wonder Soren was a slave to masturbation!
Maybe so, but I think that in the case of the perfectly playing entity today, it doesn't allow itself to be drawn into such lines.
It's kind of presumptuous for you to state that you know what a drawish line is for a perfect entity based on today's badly flawed engines. These kind of things are disproven every time a new generation of engines comes out.
These kinds of things are primarily disproven in highly tactical positions. Especially with some of the d4 openings, however, these positions tend not to arise, and the strategical evaluation is much more important. It is also in many of these types of positions that small mistakes in the evaluation tend not to be very costly for powerful engines. I think that a super-grandmaster could put together an opening book that would allow something like Hiarcs or Fritz to draw against the Rybka Cluster most of the time.
They are only disproven in lines that can be resolved to a known game state (win, lose or draw) in a short enough horizon to be solvable with current tools. Perfect play would avoid any such lines in favor of lines that require an horizon well beyond the capabilities of engines today or in the near future.
The Rybka cluster's horizon goes out only slightly farther (maybe the equivalent of a couple of plys) than the latest Hiarcs or Fritz on the latest hardware, yet it is a much better engine. This doesn't say much about a competition where one side sees all lines to their conclusions, while the other uses today's engines which achieve high plycounts only by pruning moves which in many cases are required to maintain a winning position.
The Rybka cluster's horizon goes out only slightly farther (maybe the equivalent of a couple of plys) than the latest Hiarcs or Fritz on the latest hardware, yet it is a much better engine. This doesn't say much about a competition where one side sees all lines to their conclusions, while the other uses today's engines which achieve high plycounts only by pruning moves which in many cases are required to maintain a winning position.
> It's kind of presumptuous...
This entire argument is futile in that it is based on presumptions!
There is no perfect play Elo!
Of course there is. Most of us admit that we have no idea what it is, while other claim to be able to pinpoint it to within 5 Elo. In either case, not knowing what it is doesn't mean it doesn't exist!
Taking into consideration the entity that is writing the code-the technological advances in hardware, and general knowledge- at this point in time it is beyond our comprehension. What makes you think that Elo isn't fluid- that it doesn't stagnate at any given point in time.
When we reach a point where we can definitively put each move into one of three categories, leading to a win, loss, or draw, we will have solved chess. Right now, engines try to estimate the probability of a move leading to one of those three possibilities, and play by trying to increase the probability of a win and decrease the probability of a loss. Even exponential improvements in hardware and in the strength of software (as measured in Elo) won't make much of a dent in progress toward the perfect player, if being a perfect player requires a horizon of hundreds of moves.
You cannot even concieve of chess being solved. Do you think that the advancement in ELO could be the answer to the solving of chess?!
Chess is a problem where each quantum of improvement (say 1 Elo) requires exponentially (at least) more work than the one that preceded it. Although engines have surpassed humans in this endeavor, I believe that we are still closer to the starting line than the finish. People tend to confuse the state where it is so difficult to make progress on a problem that things are almost at a standstill, with reaching the endpoint. This may be true for practical purposes, but it doesn't mean that we've actually solved the problem. It means we've taken it as far as we can go.
I claim that I'm able to pinpoint it within 4000 elo, but that my estimate will be much closer to reality than mindbreaker's
> but the non-perfect player will get worse
But what if the "non-perfect player" had a holographic concept of reality!
Not getting better with more time (or worse with less time) is a necessary, but not sufficient condition for perfect play. It is of course possible for some entity to be equally non-perfect in any time control (e.g. the random player).
> It is of course possible for some entity to be equally non-perfect in any time control (e.g. the random player).
Actually, I would tend to believe that that would be more the case than either/or (not to be confused with Soren's "stages of existence").
It is "Random" there would not be any constancy in one direction or the other. That would be the non-perfect player is the perfect player- Coincidentia oppositorum!
Perfection is to be non-perfect-it is all inclusive. An entity that thinks it is perfect or viewed as perfect has a limitation by own definition of perfection. It is exclusive and not all inclusive. It cannot draw anything into itself- anything outside of itself is imperfect- its ability to learn is limited to itself knowledge.
The 40/120 was the time control for which I estimated 3800. I don't really think about Elo on time scales like 1 move per two days.
Perfect play has to be time control invariant. It make no sense for perfect play for 1 move in two days to be even a skosh better than perfect play at classical time controls.
Sure it does--perfect play at 1 move in two days will be a move that makes things more difficult for the opponent to find a perfect response. Note the term "a" perfect response. Most moves have many perfect responses, even if things aren't made as difficult as possible on the opponent. For example, I think that the probability that my recent Wing Gambit game with Paul was a perfect game is very close to 100%, but I'm sure that if one of us had a much stronger opponent, it would have been more difficult to find the perfect moves every time. One of the things that opponent would have to do would be to go into slightly inferior, but still perfect, and much more difficult lines, and hope for a slip.
I think it is very doubtful that your wing gambit game with Paul never crossed the perfect play draw boundary.
For a perfect player, there is no such thing as a slightly inferior position. It's either won, lost, or drawn. There is however the possibility of frequently presenting the opponent with choices that require a horizon of hundreds of moves to correctly resolve, while never crossing the draw boundary. If the perfect player chooses lines that offer a large number of these choices, the non-perfect player's chance of making it to a dead draw will be extremely low.
For a perfect player, there is no such thing as a slightly inferior position. It's either won, lost, or drawn. There is however the possibility of frequently presenting the opponent with choices that require a horizon of hundreds of moves to correctly resolve, while never crossing the draw boundary. If the perfect player chooses lines that offer a large number of these choices, the non-perfect player's chance of making it to a dead draw will be extremely low.
I think that in real games, there are very, very few positions in which such a long horizon is required. The reasons that I can say so confidently that Paul and I never crossed the draw boundary are (1) if this was done, then this was done multiple times, and the probability of getting back to a draw without a very obvious "saving maneuver" is very slim--this is simply the random walk scenario, and (2) extensions of the main lines all showed very similar drawn endgames, even from only a few moves into the game. The Wing Gambit, unfortunately, is not the type of opening that one should play to avoid a draw--I think it is actually a very drawish opening at correspondence level. If we had played a B90 or B97, on the other hand, I'm sure there would have been a much higher probability of crossing the draw boundary.
You are attempting to use your hypothesis that current engines are close to perfect to prove that current engines are close to perfect. In reality, current engines, including R4, are seen to be wrong all the time during cc games. There is no way for you to know what the true state of any game is until minimax of the complete tree reaches a result of draw by stalemate, repetition, or 50-move rule, a decisive ending, or a tablebase position. Saying that you know that one of your games was perfect (never crossed the draw barrier) reflects badly on your understanding of the problem.
No, all I'm doing is using the random walk problem to show this. Human understanding of these things goes quite high, too, i.e. a decent enough human can usually tell when a normal game (i.e. not one of these theoretical unclear tablebase endings) has crossed the draw barrier, or has at least come close to doing so. We weren't sure of the true nature of the position against Shahar, for example, but we all knew we were in trouble. To cross it precisely an even number of times and end at a draw without anyone noticing is quite improbable.
Human understanding of these things goes quite high, too, i.e. a decent enough human can usually tell when a normal game (i.e. not one of these theoretical unclear tablebase endings) has crossed the draw barrier, or has at least come close to doing so.
Surely you jest! There are no shortage of complicated positions where a top GM will have no idea whether a game belongs in the win, lose, or draw column. There is another set of positions where top GMs will think they know which column the game belongs in, but will be mistaken. The idea that you and/or Shahar can pull off this feat is laughable! The reality is that you have no clue whether your game crossed the draw barrier any number of times.
I won't even get into the criteria for a "normal" game because this is really too silly to contemplate.
Surely you jest! There are no shortage of complicated positions where a top GM will have no idea whether a game belongs in the win, lose, or draw column. There is another set of positions where top GMs will think they know which column the game belongs in, but will be mistaken. The idea that you and/or Shahar can pull off this feat is laughable! The reality is that you have no clue whether your game crossed the draw barrier any number of times.
I won't even get into the criteria for a "normal" game because this is really too silly to contemplate.
Nonetheless, the more times you think that it has crossed, the more improbable it is that the game will actually end in a draw due to the random walk phenomenon. Also, if a strong player has no chance at even recognizing such positions, than the draw percentage wouldn't go up so substantially as you get toward a higher level of play.
You have made two unsupported assertions:
1) The more times you think that it has crossed, the more improbable it is that the game will actually end in a draw due to the random walk phenomenon.
2) If a strong player has no chance at even recognizing such positions, than the draw percentage wouldn't go up so substantially as you get toward a higher level of play.
Suppose two entities are playing a game and arrive at a position where there are two moves that look reasonable. One is a draw, but the other leads to mate in 1000 (with the person about to move losing of course). If the game is being played by two top GMs and the player to move cannot recognize the mate in 1000 that he will suffer from if he plays one of the moves, he may choose the losing move and cross the draw boundary. His opponent at that point may not recognize the mate in 999 and not play a winning move.
OK. Here are are my unsupported assertions:
1) There is no way to tell if my hypothesized situation occurs multiple times in a game that looks like it was drawn from start to finish (an even number of times if the starting position is a draw), and
2) Top players, both human and engine, are trained to evaluate positions based on their chances of being able to win them. If getting the win against good play requires a 1000 move sequence, it is very likely that the position will not be considered advantageous by human or engine, much as some six-man TB positions would not be sought out by humans or engines without TB access.
My conclusion is that in general, the level of play to get a win is generally going to be higher than the level of play necessary to get a draw, and this leads to more and more draws for equally matched players as their playing level increases. This is NOT in any way indicative of perfect play though.
1) The more times you think that it has crossed, the more improbable it is that the game will actually end in a draw due to the random walk phenomenon.
2) If a strong player has no chance at even recognizing such positions, than the draw percentage wouldn't go up so substantially as you get toward a higher level of play.
Suppose two entities are playing a game and arrive at a position where there are two moves that look reasonable. One is a draw, but the other leads to mate in 1000 (with the person about to move losing of course). If the game is being played by two top GMs and the player to move cannot recognize the mate in 1000 that he will suffer from if he plays one of the moves, he may choose the losing move and cross the draw boundary. His opponent at that point may not recognize the mate in 999 and not play a winning move.
OK. Here are are my unsupported assertions:
1) There is no way to tell if my hypothesized situation occurs multiple times in a game that looks like it was drawn from start to finish (an even number of times if the starting position is a draw), and
2) Top players, both human and engine, are trained to evaluate positions based on their chances of being able to win them. If getting the win against good play requires a 1000 move sequence, it is very likely that the position will not be considered advantageous by human or engine, much as some six-man TB positions would not be sought out by humans or engines without TB access.
My conclusion is that in general, the level of play to get a win is generally going to be higher than the level of play necessary to get a draw, and this leads to more and more draws for equally matched players as their playing level increases. This is NOT in any way indicative of perfect play though.
What you're saying in the mate in 1000 scenario is that players tend to play in a way that goes precisely against the random walk phenomenon. However, this cannot be true: if this was the case, then they would basically be playing in the style of a "help-mate"--except here, it is in the style of a "half-draw". Instead, players are reasonably strong: if it were just for pure randomness, then once a position has crossed the draw boundary, then it is just as likely to get back to a draw as it is to go farther from the draw, i.e. for the win to become more clear. However, there are two things that make it NOT pure randomness, and I think that both of these go against your arguments. First, if it is as you say that the players tend to go back toward the draw, then there must be far more reasonable-looking drawing moves than there are winning moves. If that's the case, then perfect play is easier than I imagined. Second, the reality is that players DO try to go for wins if they think it is possible. This will add another tendency to bring the game away from the draw if it is possible to do so. Again, if it is as you say that the players simply miss the wins, then we go back to the random walk problem (after all, they will miss the draws, too, and it will keep shifting back and forth--though much more to one side the longer the game continues due to the square root of N law), which, when invoked, proves that you cannot have very many draws. I think all of this makes your "unsupported assertions" moot, with number 2 already having been partially used as part of the random walk argument.
You've missed the point entirely. If a position exists a shortest distance to mate in 1000 moves, in the great majority of cases there will also exist mates in 1000+ moves, and many moves are likely to lead to drawn positions as well. A person or an engine that can't distinguish between these three classes (optimal mate sequence, longer/shorter mate sequence, or a move to draw (by the player who is to move and win) will likely play moves randomly from the three categories until the distance to mate or to draw is much less than 1000. At this point the game is much more likely to end in a draw if both players see this coming at about the same time.
In any event, your thesis that you, or you and Shahar, or anyone else, has any clue as to whether your game strayed into a won/lost position for one of the players is purely a guess and very possibly an incorrect one at that. Furthermore, it is very possible that both sides at some point in the game had won positions that they were totally ignorant of and blundered away into an easier to reach drawn endgame. This is not at all uncommon, and there are a fair number of won 6-man positions that will almost always end up as draws without TB support, without either player knowing that the game was won or lost.
In any event, your thesis that you, or you and Shahar, or anyone else, has any clue as to whether your game strayed into a won/lost position for one of the players is purely a guess and very possibly an incorrect one at that. Furthermore, it is very possible that both sides at some point in the game had won positions that they were totally ignorant of and blundered away into an easier to reach drawn endgame. This is not at all uncommon, and there are a fair number of won 6-man positions that will almost always end up as draws without TB support, without either player knowing that the game was won or lost.
I believe that in the cases that have been discovered so far, the really long mates have been the rare ones. If one believes that chess is actually a draw from the initial position, which seems very likely, then one would have to believe that in the positions that very high-level players actually obtain from their games, these positions with mates in 1000, much less with other mates in more than 1000, do not generally arise without the eventual realization of a possible "missed" opportunity. In the present day, the "clear wins" are positions that are probably mate in about 50 or 60 moves. Then there are others, where the position is "unclear", and these are the types of positions that people with big hardware like to play because they are confident that their hardware will "clear things up" more quickly. It is these unclear positions where you are probably often talking about mates in hundreds of moves for one side or the other...occasionally. Nonetheless, if the original position is drawn, then the probability of being able to stray significantly from that with such a complex game ahead without strong players noticing is fairly slim...otherwise, the probability that the original position is drawn is incredibly slim--like being balanced on the edge of a knife that is already leaning one way (in the direction of white, whose turn it is to move).
I believe that in the cases that have been discovered so far, the really long mates have been the rare ones.
There are a large number of long (>100 move) mates with 6 pieces. There is no reason to think that with an arbitrary number of pieces, there aren't an enormous number of long mates. The percentage relative to draws is irrelevant if at least one of the players can see the long mates. In the current case where neither player can see the long mates, the small percentage of wins is due to the players missing winning opportunities that they don't see. There is absolutely no reason to believe that this doesn't happen all the time.
Nonetheless, if the original position is drawn, then the probability of being able to stray significantly from that with such a complex game ahead without strong players noticing is fairly slim...otherwise, the probability that the original position is drawn is incredibly slim--like being balanced on the edge of a knife that is already leaning one way (in the direction of white, whose turn it is to move).
This is just nonsense. The original position is probably drawn, as are the great majority of quiet 32-man positions, but there is no assurance of this. And your assertion that a strong player can tell whether a game is drawn or not, when it is easy enough to pick out similar 6-man positions where one is won and the other is drawn and no human or engine can tell the difference without TBs, is totally indefensible.
Your knife edge analogy is also wrong. If the great majority of quite and balanced positions are drawn and only a small percentage are won, the tendency of players who can't see the wins from the draws will be to wander back into drawn positions, due to their much higher percentages. The drawn position should be the stable output for equally strong players who aren't making simple blunders, just as it is observed to be.
There are a large number of long (>100 move) mates with 6 pieces. There is no reason to think that with an arbitrary number of pieces, there aren't an enormous number of long mates. The percentage relative to draws is irrelevant if at least one of the players can see the long mates. In the current case where neither player can see the long mates, the small percentage of wins is due to the players missing winning opportunities that they don't see. There is absolutely no reason to believe that this doesn't happen all the time.
Nonetheless, if the original position is drawn, then the probability of being able to stray significantly from that with such a complex game ahead without strong players noticing is fairly slim...otherwise, the probability that the original position is drawn is incredibly slim--like being balanced on the edge of a knife that is already leaning one way (in the direction of white, whose turn it is to move).
This is just nonsense. The original position is probably drawn, as are the great majority of quiet 32-man positions, but there is no assurance of this. And your assertion that a strong player can tell whether a game is drawn or not, when it is easy enough to pick out similar 6-man positions where one is won and the other is drawn and no human or engine can tell the difference without TBs, is totally indefensible.
Your knife edge analogy is also wrong. If the great majority of quite and balanced positions are drawn and only a small percentage are won, the tendency of players who can't see the wins from the draws will be to wander back into drawn positions, due to their much higher percentages. The drawn position should be the stable output for equally strong players who aren't making simple blunders, just as it is observed to be.
But for engine "The Wall" to play the defense, would have to know how the offense was going to go, so would also have an equally strong offense! Hence, there really is no "The Wall" engine--it would crush you @ the top elo.
>how long would deepRybka 4 have to think to = Perfect ELO?
Unfortunately Rybka is inevitably pruning some best moves in some positions, so no amount of think time would result in Rybka finding it. As far as I know anyways.
The pruning in Rybka is from the Time Control right? So the longer the time control, the less pruning is the matter right? If the pruning were not of the Time Control itself, I wonder what Rybka's ELO would "top out" at from the pruning it has and at what time control dR4 would stop getting better!
![[gb]](/mwf/v12/flags/gb.png "United Kingdom")
So the longer the time control, the less pruning is the matter right?
:
No. Longer time control, same pruning, greater depth.
:
:
No. Longer time control, same pruning, greater depth.
:
![[il]](/mwf/v12/flags/il.png "Israel")
How do you know it?
Did you write the code of rybka?
Are you sure that the pruning of best move is not big reductions(when the depth is big enough) so rybka may need depth 50 and not depth 20 to find the move but eventually rybka can find it?
Did you write the code of rybka?
Are you sure that the pruning of best move is not big reductions(when the depth is big enough) so rybka may need depth 50 and not depth 20 to find the move but eventually rybka can find it?
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill