Don't be fooled by this picture, which was taken a year ago when I was 70. My mind still works, anyhow, which is more than can be said for many others here a lot younger.
Chess will never be solved, here's why
Sort:
DesperateKingWalk wrote:
Botlosenik wrote:
Optimissed wrote:
It seems we need to develop an algorithm which works well and which is 100% reliable. So what must be done first, in order that such an algorithm may be developed?
I made some comments on such an algorithm above, starting with "Strange thread indeed". In short, I believe that without a revolution (as in making the existing stuff seem like children's toys) in either computer hardware or chess theory, what you are asking for is impossible in the foreseeable future.
I am not sure if most people realize it. But such an algorithm is in full development. And has been in full development for years. And many people here have it installed on their computer.
An "algorithm" that tries to determine the outcome of any chess position just by looking at the chess position. And assessing if the position is a win, loss, or draw. And can do so with zero calculation or search. And can suggest moves in the chess game position. Along with the evaluation.
The "algorithm" can play a strong game of chess with no calculation. And can beat most human chess players, just from looking at the chess position But plays far from perfect chess.
I will give an example of the play from the the best "algorithm" we have on the planet. Playing against a weaker "algorithm". This game was played with zero search, and by only assessing the position on positional grounds.
As you will see these "algorithms" can play at a very high level. And better then 99.9% of all human players. But without searching. But playing perfect chess is not possible. As chess is ultimately a 100% tactical game.
The best "algorithms" in chess are called Neural Networks. And are produced by a computer, instead of being coded by a human.
The strongest of these Neural Networks are developed by Leela Chess Zero. And are in continues development. In a effort to created the strongest Neural Network possible.
The strongest of these Neural Networks can be downloaded here.
https://training.lczero.org/networks/?show_all=0
And newer nets and stronger nets are created every hour of the day. To continually and hopefully created stronger Neural Networks for chess.
These Neural Networks are how the modern chess engines play chess, and why the evaluations have improved so much in the last few years.
This games was played today using the "algorithms" only to play the game of chess. The "algorithms" were possessed by Leela Chess Zero, and Stockfish. And are setup by playing the game with 1 node as the time control in the chess GUI of Banksia.
I have added annotations to the game. And the Centipawn Analysis so you can see the level of play in this game.
Most human players score well above 1.00 Centipawns as their average error rate at 3 minutes a move.
Grandmaster score below .40 Centipawns on average as there error rate at 3 minutes a move.
GM Magnus Carlsen scored a error rate .12 Centipawns when scoring 50 of his best winning games at 3 minutes a move.
My son works in "digital intelligence" and is no doubt doing his bit towards developing it. Very often it's the progress made in specific issues, in many and diverse aspects of engineering, which conglomerates in such a way that a bright spark somewhere gets the credit for moving the research to a new paradigm.
I'll find out what he thinks. There are people here on this thread who have contributed to research but their time is probably past. When I was involved in computing my interest was writing very compact software which was as fast as possible. My interests were completely out of step with the then obsession, in the early 90s, with writing top-down, exchangeable, modular programmes that everyone could understand.
DesperateKingWalk wrote:
Botlosenik wrote:
Optimissed wrote:
It seems we need to develop an algorithm which works well and which is 100% reliable. So what must be done first, in order that such an algorithm may be developed?
I made some comments on such an algorithm above, starting with "Strange thread indeed". In short, I believe that without a revolution (as in making the existing stuff seem like children's toys) in either computer hardware or chess theory, what you are asking for is impossible in the foreseeable future.
I am not sure if most people realize it. But such an algorithm is in full development. And has been in full development for years. And many people here have it installed on their computer.
An "algorithm" that tries to determine the outcome of any chess position just by looking at the chess position. And assessing if the position is a win, loss, or draw. And can do so with zero calculation or search. And can suggest moves in the chess game position. Along with the evaluation.
The "algorithm" can play a strong game of chess with no calculation. And can beat most human chess players, just from looking at the chess position But plays far from perfect chess.
I will give an example of the play from the the best "algorithm" we have on the planet. Playing against a weaker "algorithm". This game was played with zero search, and by only assessing the position on positional grounds.
As you will see these "algorithms" can play at a very high level. And better then 99.9% of all human players. But without searching. But playing perfect chess is not possible. As chess is ultimately a 100% tactical game.
The best "algorithms" in chess are called Neural Networks. And are produced by a computer, instead of being coded by a human.
The strongest of these Neural Networks are developed by Leela Chess Zero. And are in continues development. In a effort to created the strongest Neural Network possible.
The strongest of these Neural Networks can be downloaded here.
https://training.lczero.org/networks/?show_all=0
And newer nets and stronger nets are created every hour of the day. To continually and hopefully created stronger Neural Networks for chess.
These Neural Networks are how the modern chess engines play chess, and why the evaluations have improved so much in the last few years.
This games was played today using the "algorithms" only to play the game of chess. The "algorithms" were possessed by Leela Chess Zero, and Stockfish. And are setup by playing the game with 1 node as the time control in the chess GUI of Banksia.
I have added annotations to the game. And the Centipawn Analysis so you can see the level of play in this game.
Most human players score well above 1.00 Centipawns as their average error rate at 3 minutes a move.
Grandmaster score below .40 Centipawns on average as there error rate at 3 minutes a move.
GM Magnus Carlsen scored a error rate of .12 Centipawns when scoring 50 of his best winning games at 3 minutes a move.
I am well aware of Leela and Alpha Zero chess, including their algorithms. As long as I followed the battle between Leela and Stockfish, it wasn't very clear which was the best, so the AZ/Leela approach is not yet established as "the only way", at least not yet. And it may be harder to improve on Leela than Stockfish actually, time will tell. Yes, Leela can train against itself with no human interaction, but there could be a upper limit to how well that works. Then on the other hand, if you want to use your rig for other games, Leela/Alpha Zero use the same software for many games, with self training, which means that it can get "as good at another game as it is at chess" without extra work for humans, so the approach has clear advantages. Alpha Zero for example was also trained to play Go (that was the first game) and chinese chess I think it was (too lazy to look it up). Oh, and by the way, the neural networks indeed get really good, "superhuman", at "intuition" about a position as you say, most people will be crushed by the neural network alone, but the programs, alpha zero and leela, do not only use that, they also use a type of tree search.
Either way, the leela/alpha zero approach as of now at least, does not lend itself well to the attempt to solve a game, only to train itself to become a very strong player.
That seems right if only from the comments of others who don't understand that but who point to Leela etc in an attempt to contradict a perfectly good observation.
Dec 31, 2022
0
#6866
Optimissed wrote:
Do you know the difference between being a clever-clogs 13 year old and the mature, wise and reflective 16 year old you should aspire to emulate? Apparently not. Grow up and I'll treat you as an adult.
Otherwise, don't tell others what they can know and what they can think when your own cognitive faculties are extremely under-developed.
Thank you. Happy new year to you too!
Dec 31, 2022
0
#6867
DesperateKingWalk wrote:
Botlosenik wrote:
DesperateKingWalk wrote:
Botlosenik wrote:
Optimissed wrote:
It seems we need to develop an algorithm which works well and which is 100% reliable. So what must be done first, in order that such an algorithm may be developed?
I made some comments on such an algorithm above, starting with "Strange thread indeed". In short, I believe that without a revolution (as in making the existing stuff seem like children's toys) in either computer hardware or chess theory, what you are asking for is impossible in the foreseeable future.
I am not sure if most people realize it. But such an algorithm is in full development. And has been in full development for years. And many people here have it installed on their computer.
An "algorithm" that tries to determine the outcome of any chess position just by looking at the chess position. And assessing if the position is a win, loss, or draw. And can do so with zero calculation or search. And can suggest moves in the chess game position. Along with the evaluation.
The "algorithm" can play a strong game of chess with no calculation. And can beat most human chess players, just from looking at the chess position But plays far from perfect chess.
I will give an example of the play from the the best "algorithm" we have on the planet. Playing against a weaker "algorithm". This game was played with zero search, and by only assessing the position on positional grounds.
As you will see these "algorithms" can play at a very high level. And better then 99.9% of all human players. But without searching. But playing perfect chess is not possible. As chess is ultimately a 100% tactical game.
The best "algorithms" in chess are called Neural Networks. And are produced by a computer, instead of being coded by a human.
The strongest of these Neural Networks are developed by Leela Chess Zero. And are in continues development. In a effort to created the strongest Neural Network possible.
The strongest of these Neural Networks can be downloaded here.
https://training.lczero.org/networks/?show_all=0
And newer nets and stronger nets are created every hour of the day. To continually and hopefully created stronger Neural Networks for chess.
These Neural Networks are how the modern chess engines play chess, and why the evaluations have improved so much in the last few years.
This games was played today using the "algorithms" only to play the game of chess. The "algorithms" were possessed by Leela Chess Zero, and Stockfish. And are setup by playing the game with 1 node as the time control in the chess GUI of Banksia.
I have added annotations to the game. And the Centipawn Analysis so you can see the level of play in this game.
Most human players score well above 1.00 Centipawns as their average error rate at 3 minutes a move.
Grandmaster score below .40 Centipawns on average as there error rate at 3 minutes a move.
GM Magnus Carlsen scored a error rate of .12 Centipawns when scoring 50 of his best winning games at 3 minutes a move.
I am well aware of Leela and Alpha Zero chess, including their algorithms. As long as I followed the battle between Leela and Stockfish, it wasn't very clear which was the best, so the AZ/Leela approach is not yet established as "the only way", at least not yet. And it may be harder to improve on Leela than Stockfish actually, time will tell. Yes, Leela can train against itself with no human interaction, but there could be a upper limit to how well that works. Then on the other hand, if you want to use your rig for other games, Leela/Alpha Zero use the same software for many games, with self training, which means that it can get "as good at another game as it is at chess" without extra work for humans, so the approach has clear advantages. Alpha Zero for example was also trained to play Go (that was the first game) and chinese chess I think it was (too lazy to look it up). Oh, and by the way, the neural networks indeed get really good, "superhuman", at "intuition" about a position as you say, most people will be crushed by the neural network alone, but the programs, alpha zero and leela, do not only use that, they also use a type of tree search.
Either way, the leela/alpha zero approach as of now at least, does not lend itself well to the attempt to solve a game, only to train itself to become a very strong player.
Correct. But we are talking about the network only. Without search. And as I said the game was played with zero search.
And as of today. Lc0 has the strongest stand alone no search chess play.
But Chess engines have to do both evalution, and search for strongest overall chess play.
In this regard as a chess playing chess engine with search, or look ahead.
Stockfish is the strongest chess player on this planet. As Stockfish has a better search then Lc0, but a weaker Neural Network.
Remember because of the size of the Leela Chess Zero network and Bigger is better most often with Neural Networks. Leela Chess Zero must be played using a GPU for strongest play. This is because of the size of the Leela Chess Zero's Neural Network. Nothing is for free. Bigger Neural Network, the slower the search aspect of the chess engine.
NNUE a smaller neural network was developed for Stockfish. And tries to get the best of both worlds. A benefit from the smaller but dumber neural network, but as a result can search much deeper in the chess position.
As of today. Stockfish NNUE approach is better then the Lc0 approach. Faster search, smaller and dumber net vs Big smart net, slower search.
But @Botlosenik is obviously correct. Neither SF nor Leela are relevant to solving chess; only to playing strongly against other limited lookahead players.
DesperateKingWalk wrote:
MARattigan wrote:
DesperateKingWalk wrote:
Botlosenik wrote:
DesperateKingWalk wrote:
Botlosenik wrote:
Optimissed wrote:
It seems we need to develop an algorithm which works well and which is 100% reliable. So what must be done first, in order that such an algorithm may be developed?
I made some comments on such an algorithm above, starting with "Strange thread indeed". In short, I believe that without a revolution (as in making the existing stuff seem like children's toys) in either computer hardware or chess theory, what you are asking for is impossible in the foreseeable future.
I am not sure if most people realize it. But such an algorithm is in full development. And has been in full development for years. And many people here have it installed on their computer.
An "algorithm" that tries to determine the outcome of any chess position just by looking at the chess position. And assessing if the position is a win, loss, or draw. And can do so with zero calculation or search. And can suggest moves in the chess game position. Along with the evaluation.
The "algorithm" can play a strong game of chess with no calculation. And can beat most human chess players, just from looking at the chess position But plays far from perfect chess.
I will give an example of the play from the the best "algorithm" we have on the planet. Playing against a weaker "algorithm". This game was played with zero search, and by only assessing the position on positional grounds.
As you will see these "algorithms" can play at a very high level. And better then 99.9% of all human players. But without searching. But playing perfect chess is not possible. As chess is ultimately a 100% tactical game.
The best "algorithms" in chess are called Neural Networks. And are produced by a computer, instead of being coded by a human.
The strongest of these Neural Networks are developed by Leela Chess Zero. And are in continues development. In a effort to created the strongest Neural Network possible.
The strongest of these Neural Networks can be downloaded here.
https://training.lczero.org/networks/?show_all=0
And newer nets and stronger nets are created every hour of the day. To continually and hopefully created stronger Neural Networks for chess.
These Neural Networks are how the modern chess engines play chess, and why the evaluations have improved so much in the last few years.
This games was played today using the "algorithms" only to play the game of chess. The "algorithms" were possessed by Leela Chess Zero, and Stockfish. And are setup by playing the game with 1 node as the time control in the chess GUI of Banksia.
I have added annotations to the game. And the Centipawn Analysis so you can see the level of play in this game.
Most human players score well above 1.00 Centipawns as their average error rate at 3 minutes a move.
Grandmaster score below .40 Centipawns on average as there error rate at 3 minutes a move.
GM Magnus Carlsen scored a error rate of .12 Centipawns when scoring 50 of his best winning games at 3 minutes a move.
I am well aware of Leela and Alpha Zero chess, including their algorithms. As long as I followed the battle between Leela and Stockfish, it wasn't very clear which was the best, so the AZ/Leela approach is not yet established as "the only way", at least not yet. And it may be harder to improve on Leela than Stockfish actually, time will tell. Yes, Leela can train against itself with no human interaction, but there could be a upper limit to how well that works. Then on the other hand, if you want to use your rig for other games, Leela/Alpha Zero use the same software for many games, with self training, which means that it can get "as good at another game as it is at chess" without extra work for humans, so the approach has clear advantages. Alpha Zero for example was also trained to play Go (that was the first game) and chinese chess I think it was (too lazy to look it up). Oh, and by the way, the neural networks indeed get really good, "superhuman", at "intuition" about a position as you say, most people will be crushed by the neural network alone, but the programs, alpha zero and leela, do not only use that, they also use a type of tree search.
Either way, the leela/alpha zero approach as of now at least, does not lend itself well to the attempt to solve a game, only to train itself to become a very strong player.
Correct. But we are talking about the network only. Without search. And as I said the game was played with zero search.
And as of today. Lc0 has the strongest stand alone no search chess play.
But Chess engines have to do both evalution, and search for strongest overall chess play.
In this regard as a chess playing chess engine with search, or look ahead.
Stockfish is the strongest chess player on this planet. As Stockfish has a better search then Lc0, but a weaker Neural Network.
Remember because of the size of the Leela Chess Zero network and Bigger is better most often with Neural Networks. Leela Chess Zero must be played using a GPU for strongest play. This is because of the size of the Leela Chess Zero's Neural Network. Nothing is for free. Bigger Neural Network, the slower the search aspect of the chess engine.
NNUE a smaller neural network was developed for Stockfish. And tries to get the best of both worlds. A benefit from the smaller but dumber neural network, but as a result can search much deeper in the chess position.
As of today. Stockfish NNUE approach is better then the Lc0 approach. Faster search, smaller and dumber net vs Big smart net, slower search.
But @Botlosenik is obviously correct. Neither SF nor Leela are relevant to solving chess; only to playing strongly against other limited lookahead players.
That is correct. As I explained here before. Stockfish and the other chess engines. Are type B Shannon chess engines. Only designed to defeat human players.
Type B Shannon chess engines like Stockfish. Were not designed to solve chess. And could not solve chess even on a computer with infinite speed, and time. Because it was not designed to do such a task.
Hm, if you throw in infinite memory in addition, then Stockfish algorithm with search depth tweaked to something like 10000 ply (max length of a legal chessgame) can trivially play perfect games, and also in analysis mode trivially tell you which moves, in any position are winning/drawing/losing. Sounds like "solve" to me.
@7018
"then Stockfish algorithm with search depth tweaked to something like 10000 ply"
++ There is no need or even use for such a depth. Depth 200 ply is more than enough.
"play perfect games, and also in analysis mode trivially tell you which moves, in any position are winning/drawing/losing." ++ In any position would be strongly solving and that is not feasible as 10^44 legal positions are too much. However, all paths from the initial position to a 7-men endgame table base draw is doable in 5 years and weakly solves chess.
Here is an example:
https://www.iccf.com/game?id=1164344
A perfect game in 57 moves from the initial position to a 7-men table base draw with optimal play from both sides.
@6988
"Without a definition, in fact, even with a definition, even only one assistant will give different answers at different times. Humans just aren't as consistent as you believe."
++ An ICCF (grand)master will not resign in a drawn position or agree on a draw in a won position. The human intervention is the equivalent of either resigning or agreeing on a draw. After 1 e4 e5 2 Ba6? white resigns. The question is not about evaluating a position, that cannot be done neither by a human, not by an engine, only by calculating towards the 7-men endgame table base. The question is about dismissing certain legal possibilities like 1 e4 e5 2 Ba6? or about terminating calculation like in a drawn opposite color bishop endgame.
"It could very well be that starting pos is a draw, and after Ba6 it is still draw."
++ No. The initial position is a draw. The position after 1 e4 e5 2 Ba6? is a loss for white.
I have even given a proof game above.
"So most of the time the assessment of the assistant is unnecessary"
++ The assistants do not assess positions. They launch the calculation from meaningful starting points (not from 1 e4 e5 2 Ba6?) and they occasionally terminate a calculation when they know for sure a draw is inevitable, like in opposite colored bishop endgames.
"all of the time it is an incredible slowdown" ++ No, the humans speed up the process by not launching pointless calculations like from 1 e4 e5 2 Ba6? and by occasionally terminating pointless calculations like in drawn opposite colored bishop endgames.
"Deep Blue was capable of evaluating 200 million positions per second"
++ Modern cloud engines reach 1 billion nodes per second.
"You want to put a human in there to evaluate every position where there is a clear material difference." ++ No. The human launches the calculation from a meaningful starting point, preferably with 26 men and either equal material, or unequal material with identifyable compensation. Then the engine calculates until the 7-men endgame table base essentially without any human intervention. In rare cases the human intervenes to terminate a calculation when the outcome of a draw is inevitable like in opposite colored bishop endgames. It is the humans from 32 to 26 men, the engine from 26 to 7 men, and the table base at 7 men.
"Lets say something like e4 a6 Bxa6 Nxa6"
++ That is a clear loss for white. No need to launch any calculation from there.
1 e4 a6 might be enough to draw or not, but it would not be the first choice of the assistants, that would be 1 e4 e5. So 1 e4 a6 would not be calculated at all.
"I hope you are aware that computers typically do selective search already, so they do spend less time on weak moves?" ++ Once a calculation is launched, the humans do not intervene, except in rare cases like after reaching a drawn opposite colored bishop endgame.
"Or the Alpha Zero algorithm, which will use neural nets to do super fast and often superior guesses of the value of a position without the need for a human?"
++ The point is that in solving Chess, for most positions we cannot depend on any evaluation, either human, or AlphaZero algorithm, as all such evaluations are known to be wrong once in a while. The only way is to calculate with enough depth to reach the 7-men endgame table base.
That is also what GM Sveshnikov said: "I will bring all openings to technical endgames".
So start from a meaningful opening (not 1 e4 e5 2 Ba6?) and then calculate until the 7-men endgame table base.
tygxc wrote:
@7018
"then Stockfish algorithm with search depth tweaked to something like 10000 ply"
++ There is no need or even use for such a depth. Depth 200 ply is more than enough.
"play perfect games, and also in analysis mode trivially tell you which moves, in any position are winning/drawing/losing." ++ In any position would be strongly solving and that is not feasible as 10^44 legal positions are too much. However, all paths from the initial position to a 7-men endgame table base draw is doable in 5 years and weakly solves chess.
Here is an example:
https://www.iccf.com/game?id=1164344
A perfect game in 57 moves from the initial position to a 7-men table base draw with optimal play from both sides.
You seem not to have taken the time to read the conversation properly before replying. What we were talking about here was INFINITE speed/memory. 10^44 is nothing compared to infinity.
Jan 1, 2023
0
#6872
btickler wrote:
...It's New Year's Eve. Time to move on from your unbreakable windmill.
Brian, it's New Year's Eve. Time to move on from your idea that anyone cares what you have to say
@7021
"10^44 is nothing compared to infinity"
++ That is right. With infinite speed chess is strongly solved in no time.
With present finite speed 10^9 nodes/s it takes 5 full years to weakly solve Chess, like done for Checkers. Weakly solving Chess takes 1000 times more than Schaeffer did for Checkers: 10^17 instead of 10^14 relevant positions.
RemovedUsername333 wrote:
Brian, it's New Year's Eve. Time to move on from your idea that anyone cares what you have to say
Are you still around? You've been so quiet I thought you must have dropped out of high school or something. Still going with the bad info somebody gave you, I see.
P.S. Pretty childish, doing all those downvotes with your alt accounts. That's 3 in as many minutes. Whereas the downvotes you countered with upvotes on your own post came from different people.
P.P.S. "Candor"
Jan 1, 2023
0
#6875
Botlosenik wrote:
tygxc wrote:
@7018
"then Stockfish algorithm with search depth tweaked to something like 10000 ply"
++ There is no need or even use for such a depth. Depth 200 ply is more than enough.
"play perfect games, and also in analysis mode trivially tell you which moves, in any position are winning/drawing/losing." ++ In any position would be strongly solving and that is not feasible as 10^44 legal positions are too much. However, all paths from the initial position to a 7-men endgame table base draw is doable in 5 years and weakly solves chess.
Here is an example:
https://www.iccf.com/game?id=1164344
A perfect game in 57 moves from the initial position to a 7-men table base draw with optimal play from both sides.
You seem not to have taken the time to read the conversation properly before replying. What we were talking about here was INFINITE speed/memory. 10^44 is nothing compared to infinity.
This is a fair point to make in all cander. One property of infinity is that it is always greater than any finite number. This can be demonstrated through a simple example. Imagine a counting sequence starting at 1 and continuing indefinitely. No matter what finite number you choose, there will always be a larger number in the sequence. For example, if you choose the number 10, there will be 11, 12, 13, and so on in the sequence. If you choose the number 100, there will still be larger numbers in the sequence such as 101, 102, 103, and so on. This illustrates that no matter how large a finite number you choose, there will always be a larger number in the sequence, and therefore infinity is larger than any finite number.
This idea can be extended to the concept of infinite speed or memory. If something were to have infinite speed or memory, it would be able to perform tasks or store information at a rate that is limitless and unbounded. In comparison, a finite speed or memory, no matter how large it may be, would always be limited and bounded in some way. For example, a computer with 10^44 bytes of memory would eventually fill up and be unable to store any more information, whereas a computer with infinite memory would never reach this limit.
@7026
"what that means in ever solving the game of chess with Stockfish"
++ Just give enough time to reach enough depth to hit the 7-men endgame table base.
Jan 1, 2023
0
#6877
Stockfish set to 10000 ply would not only be geologically slow, it would also be unreliable. It is obvious it can only get to any large depth by ignoring lines that a cursory evaluation concludes are "not promising". Such evaluations are wrong at a rate that is not only non-zero but significant to getting the right answer (hence Stockfish' unquestionable imperfection).
Note that an engine losing a game means it got the choice of a move dead wrong at some point. What it thought was a move that was the very best was actually a blunder that gave away the game. That sentence shows the degree to which engines are unreliable. This doesn't stop being true with more time per move, it just becomes rarer (at very high computational cost).
@7028
"Stockfish set to 10000 ply"
++ There is no need for 10,000 ply. 238 ply is more than enough.
Perfect games from the ICCF WC finals lasted between 13 and 119 moves, 42 moves average, with standard deviation 16.
No no, please stay. You and I might have an average mental age of 16+ which puts us well ahead of the opposition. I'm 71 actually but I managed to hold the mental age bit at 17 until recently.