@6047
"You are not at all good at taking on board important points like that."
++ What points?
I am sure 10^17 is a good estimate for the number of positions relevant to weakly solving chess.
I am also sure 1 e4 e5 2 Ba6? loses for white.
I am also sure 1 a4 cannot be better than 1 e4 or 1 d4.
I am also sure 1 Nh3 cannot be better than 1 Nf3.
Chess will never be solved, here's why
Sort:
Oct 11, 2022
0
#4603
Elroch wrote:
snoozyman wrote:
Whatever the number of games of chess, it's still a BIG number, comparable to 52 factorial.
Rather amusing point about the video is that it goes wrong almost immediately. First it says that if you have a room of 23 people, there is about 50% chance of two people having the same birthday. This is correct. Then it says, "so if you walk into a room with 23 people in it, the probability of someone having the same birthday as you is 50%".
No!
Specifying you as one of the two people leaves only 23 possible pairs that could match rather than the (23 * 22 / 2) pairs that the correct fact refers to. The chance of one of the 23 people having the same birthday as you is about 1 - (364/365)^23 ~= 6% (ignoring leap years).
Ah, but @tygxc could reduce the number of sensible relevant birthdays to 3 bringing your probability of a match up to 94,134,790,219 / 94,143,178,827 (=50% for all practical purposes).
Elroch wrote:
snoozyman wrote:
Whatever the number of games of chess, it's still a BIG number, comparable to 52 factorial.
Rather amusing point about the video is that it goes wrong almost immediately. First it says that if you have a room of 23 people, there is about 50% chance of two people having the same birthday. This is correct. Then it says, "so if you walk into a room with 23 people in it, the probability of someone having the same birthday as you is 50%".
No!
Specifying you as one of the two people leaves only 23 possible pairs that could match rather than the (23 * 22 / 2) pairs that the correct fact refers to. The chance of one of the 23 people having the same birthday as you is about 1 - (364/365)^23 ~= 6% (ignoring leap years).
Lol, I had the exact same thought/reaction. That's a Tygxc-level blunder in thinking right there.
And this is the problem with such stories...they lose accuracy in the translation, because somebody lacking in expertise incorrectly equates two things that are not the same at all.
btickler wrote:
Lol, I had the exact same thought/reaction. That's a Tygxc-level blunder in thinking right there.
Just for accuracy, I don't believe that was tygxc's video. I'm pretty sure he'd know better than that.
mikekalish wrote:
Just for accuracy, I don't believe that was tygxc's video. I'm pretty sure he'd know better than that.
I didn't say it was. I said it shows the kind of thinking that characterizes Tygxc's mistaken assumptions and extrapolations 
. He takes a value, doesn't realize the parameters of what that value actually represents, then uses that value in some equation where he conflates it with some other value that actually conforms to a different set of criteria.
TL;DR version: he likes to do apple math with lots of oranges.
I figured that's what you meant....but my comment was intended to keep others from misinterpreting.
mikekalish wrote:
I figured that's what you meant....but my comment was intended to keep others from misinterpreting.
I also like to head off misinterpretations at the pass, so I understand the impulse.
@6031
"my comment was intended to keep others from misinterpreting"
++ Others misinterprete all the time.
They doubt what is right and are sure about the false.
I present facts & figures and back them up.
Oct 12, 2022
0
#4610
tygxc wrote:
@6031
"my comment was intended to keep others from misinterpreting"
++ Others misinterprete all the time.
They doubt what is right and are sure about the false.
I present facts & figures and back them up.
So why don't you back up your claim of being able to determine the game theoretic result of a position and the number of errors in games played from the position by applying your calculations to the games here and back up your claim of 1 error in 10^20 moves at 17 secs. per move on your 10^9 nps machine (which curiously didn't change when you dropped the time from 60 hours to 17 seconds).
No reason to restrict yourself to just draws or just KRPP vs. KRP positions. After the first single error you're bound to have both wins and draws in the game, so that's a pathetic excuse.
And contrary to what you keep posting as you try to wriggle out of the exercise, one of the positions is a drawn KRPP vs. KRP position.
@6036
"1 error in 10^20 moves at 17 secs. per move on your 10^9 nps machine
(which curiously didn't change when you dropped the time from 60 hours to 17 seconds)."
++ 17 s on the 10^9 NPS engine corresponds to 60 h on the engine of the paper.
"No reason to restrict yourself to just draws" ++ Yes: only drawn positions are relevant to weakly solving chess: hopping from the initial drawn position to other drawn positions until a 7-men endgame table base draw or a prior 3-fold repetition.
"or just KRPP vs. KRP positions"
++ Rook endings occur most. Rook endings can be draws despite one pawn down.
A rook ending is a major way for black to achieve the game-theoretic value of the draw.
"one of the positions is a drawn KRPP vs. KRP position anyway."
++ No, see top right in your image: White is winning DTZ 101.
It is a draw only by the 50-moves rule. Black can draw from the initial position without the 50-moves rule, so this position cannot be reached from optimal play by both sides.
Oct 12, 2022
0
#4612
tygxc wrote:
@6036
"1 error in 10^20 moves at 17 secs. per move on your 10^9 nps machine
(which curiously didn't change when you dropped the time from 60 hours to 17 seconds)."
++ 17 s on the 10^9 NPS engine corresponds to 60 h on the engine of the paper.
The engine AZ in the paper (which is not the one you propose to use) was presumably run on a Google's TPUs. From where did you get the equivalence?
"No reason to restrict yourself to just draws" ++ Yes: only drawn positions are relevant to weakly solving chess: hopping from the initial drawn position to other drawn positions until a 7-men endgame table base draw or a prior 3-fold repetition.
I'm not talking about weakly solving chess - you haven't posted a sensible description of that yet. I'm just talking about your calculations of game-theoretic result and error rates extracted from games played from a position. If they work they should work from all positions.
"or just KRPP vs. KRP positions"
++ Rook endings occur most. Rook endings can be draws despite one pawn down.
A rook ending is a major way for black to achieve the game-theoretic value of the draw.
How is any of that relevant to your calculation?
"one of the positions is a drawn KRPP vs. KRP position anyway."
++ No, see top right in your image: White is winning DTZ 101.
It is a draw only by the 50-moves rule. Black can draw from the initial position without the 50-moves rule, so this position cannot be reached from optimal play by both sides.
The top right of the image gives the game-theoretic result under basic rules.
The games were played with the 50 move rule in force as were the games on which you base your calculations of the error rates you claim as is the game you offer to solve with SF15 which is also designed to play with the 50 move rule in effect. The highlighted text in my image gives the game-theoretic result with the 50 move rule in effect. It's a draw.
The game played at 2048 secs per ply (terminated by adjudication when a six man position was reached) contains two errors under competition rules and only one under basic rules (which is neither of the former). Which type of error rate are you claiming to calculate?
Optimal play has nothing to do with your calculation. It's the validity of the calculation I'm asking you to check. How can you presume to "calculate" error rates on the assumption that there are no errors?
Why don't you stop wriggling and post your calculations for the games here? Then we can stop all this pointless discussion about your proposal to solve chess in five years.
"The engine AZ in the paper was presumably run on a Google's TPUs. From where did you get the equivalence?" ++ In the previous discussion somebody posted the NPS of the engine of the paper. So I converted to the 10^9 NPS of the current top engines.
"I'm not talking about weakly solving chess" ++ I am, that is the topic of this thread.
"you haven't posted a sensible description" ++ Yes several times. Take ICCF WC draws.
Analyse 3 alternatives per white move until a table base draw or a prior 3-fold repetition.
"calculations of game-theoretic result and error rates extracted from games"
++ The game theoretic result and the error rate are derived from ICCF WC draws. 136 games of which 9 decisive. It is impossible to fit a Poisson distribution with 127 / 136 probability of an odd number of errors / game, thus the game-theoretic value of Chess must be a draw.
Fitting a Poisson distribution with 9 / 126 probability of an odd number of errors / game is possible and leads to the error rate of 7 errors per 100 games.
"If they work they should work from all positions."
++ No, there are known anomalies like KNN vs. K, KB + wrong R-pawn vs. K
"How is any of that relevant to your calculation?"
++ Rook endings will occur most and are most relevant to weakly solving chess.
"one of the positions is a drawn KRPP vs. KRP position anyway."
++ No, see top right in your image: White is winning DTZ 101.
It is a draw only by the 50-moves rule. Black can draw from the initial position without the 50-moves rule, so this position cannot be reached from optimal play by both sides.
"as were the games on which you base your calculations of the error rates you claim as is"
++ In none of these games was the 50-moves rule triggered.
So it does not matter if it was in effect or not.
"the game you offer to solve with SF15 which is also designed to play with the 50 move rule"
++ The rule may be in effect, but does not trigger, i.e. it is the same as if it is not in effect.
"It's the validity of the calculation I'm asking you to check."
++ So what drawn KRPP vs. KRP position do you propose?
"How can you presume to "calculate" error rates on the assumption that there are no errors?"
++ I do not assume that there are no errors. On the contrary, I calculate from the known errors.
In AlphaZero autoplay there are 2.1 errors in 100 games @ 1 min / move.
In ICCF WC there are 9 errors in 136 games.
"Why don't you stop wriggling and post your calculations?"
++ Unlike others I have posted several calculations:
- Top down estimation of the number of relevant positions: legal - sensible - reachable - relevant
- Bottom up estimation of the number of relevant positions: upper bound, lower bound, estimate
- Calculation of width from AlphaZero autoplay.
- Calculation of depth from ICCF WC.
- Calculation of game-theoretic value, error / game from ICCF WC.
"proposal to solve chess in five years"
++ The solvability in 5 years results from the calculated 10^17 relevant positions.
I love this.
Oct 12, 2022
0
#4616
tygxc wrote:
"The engine AZ in the paper was presumably run on a Google's TPUs. From where did you get the equivalence?" ++ In the previous discussion somebody posted the NPS of the engine of the paper. So I converted to the 10^9 NPS of the current top engines.
Somebody posted where? I must have missed that. (Was it a peer reviewed post?)
Do you expect SF15 NPS to be the same as AZ NPS?
No harm done. The games I posted here all used SF15 and were run on an Intel(R) Pentium(R) CPU J3710 @ 1.60GHz with 2GB hash storage, so you can try your calculations on those.
"I'm not talking about weakly solving chess" ++ I am, that is the topic of this thread.
But you have previously talked about your calculations to determine the game theoretic value of the starting position and the expected error rates in SF15, which you say relate to the feasibility of your proposals. Why the sudden coyness?
"you haven't posted a sensible description" ++ Yes several times. Take ICCF WC draws.
Analyse 3 alternatives per white move until a table base draw or a prior 3-fold repetition.
Well, if you post your calculations we needn't go into that.
"calculations of game-theoretic result and error rates extracted from games"
++ The game theoretic result and the error rate are derived from ICCF WC draws. 136 games of which 9 decisive. It is impossible to fit a Poisson distribution with 127 / 136 probability of an odd number of errors / game, thus the game theoretic calue of Chess must be a draw.
Fitting a Poisson distribution with 9 / 126 probability of an odd number of errors / game is possible and leads to the error rate of 7 errors per 100 games.
Some detail missing, but you can supply full details when you apply them to the games here. That will give you the advantage of knowing that the games you base your calculations on are played by the vehicle you intend to use and you have the exact values for the time used.
"If they work they should work from all positions."
++ No, there are known anomalies like KNN vs. K, KB + wrong R-pawn vs. K
So you're now saying your calculations don't work even before you try them out on my games. Would the anomalies include KNN vs KP, KRBN vs KQR and KRPP vs KRP as well? If not you could try them on my games here and see if they work in general.
"How is any of that relevant to your calculation?"
++ Rook endings will occur most and are most relevant to weakly solving chess.
I ask again; how is that relevant to your calculation?
"one of the positions is a drawn KRPP vs. KRP position anyway."
++ No, see top right in your image: White is winning DTZ 101.
It is a draw only by the 50-moves rule. Black can draw from the initial position without the 50-moves rule, so this position cannot be reached from optimal play by both sides.
And you're considering games with the 50 move rule in effect so it's a draw.
"as were the games on which you base your calculations of the error rates you claim as is"
++ In none of these games was the 50-moves rule triggered.
So it does not matter if it was in effect or not.
The 50 move rule was triggered in some of the games I posted.
In ICCF games the players prefer to agree a draw or play into a triple repetition because they get more time in the pub. Totally irrelevant to your calculations.
"the game you offer to solve with SF15 which is also designed to play with the 50 move rule"
++ The rule may be in effect, but does not trigger, i.e. it is the same as if it is not in effect.
See last comment.
"It's the validity of the calculation I'm asking you to check."
++ So what drawn KRPP vs. KRP position do you propose?
I don't propose a drawn KRPP vs KRP position in particular. If your calculations work they should work irrespective of material or theoretical result.
I'm suggesting you try your calculations on each of the sets of games I posted, not just the games from the drawn KRPP vs KRP position I posted.
"How can you presume to "calculate" error rates on the assumption that there are no errors?"
++ I do not assume that there are no errors. On the contrary, I calculate from the known errors.
You said, "Black can draw from the initial position without the 50-moves rule, so this position cannot be reached from optimal play by both sides". Optimal play by both sides I would take to mean no errors.
In AlphaZero autoplay there are 2.1 errors in 100 games @ 1 min / move.
In ICCF WC there are 9 errors in 136 games.
"Why don't you stop wriggling and post your calculations?"
++ Unlike others I have posted several calculations:
- Top down estimation of the number of relevant positions: legal - sensible - reachable - relevant
- Bottom up estimation of the number of relevant positions: upper bound, lower bound, estimate
- Calculation of width from AlphaZero autoplay.
- Calculation of depth from ICCF WC.
- Calculation of game-theoretic value, error / game from ICCF WC.
"proposal to solve chess in five years"
++ The solvability in 5 years results from the calculated 10^17 relevant positions.
Just irrelevant crap so I shrunk it. If you post the calculations I requested we don't need to discuss it.
I suggested you stop wriggling and post the requested calculations, instead of which you've wriggled for another nine inches and not posted any calculations. I reiterate my suggestion.
You might also like to state whether the calculations relate to basic rules errors or competition rules errors since you ignored my question in your response.
@6042
"Somebody posted where? I must have missed that."
++ Many posts higher here. The engine of the paper is so much slower that the top engine.
"Do you expect SF15 NPS to be the same as AZ NPS? "
++ Thin nodes (SF) are better than thick nodes (AZ) to hit the 7-men endgame table base.
"The games I posted here..." ++ are not relevant.
"But you have previously talked about your calculations to determine the game theoretic value of the starting position and the expected error rates in SF15 which you say relate to the feasibility of your proposals." ++ Yes, both are relevant.
"if you post your calculations" ++ I have posted and reposted calculations several times.
"the games you base your calculations on are played by the vehicle you intend to use"
++ I have no 10^9 nodes / s cloud engine at my disposal.
"you have the exact values for the time used" ++ It can be extrapolated.
"you try them out on my games" ++ I have posted above 2 of your irrelevant games and pointed out you have a problem with your Stockfish version.
"Would the anomalies include KNN vs KP" ++ Yes, because of KNN vs. K
"KRBN vs KQR" ++ No
"KRPP vs KRP" ++ No
"try them on my games here" ++ I have posted above on some of your games and pointed out you have a problem with your Stockfish version.
"++ Rook endings will occur most and are most relevant to weakly solving chess.
I ask again; how is that relevant to your calculation?"
++ Because KRPP vs. KRP draws turn up frequently in weakly solving Chess.
"you're considering games with the 50 move rule in effect so it's a draw."
++ It is a position that does not show up in solving Chess.
Most of the ICCF WC draws >99% sure to be perfect games do not even reach 50 moves.
In none of the ICCF WC games is the 50-moves rule triggered.
7-men table base wins can be claimed if they exceed 50 moves without pawn move or capture, but such claims do not happen.
"The 50 move rule was triggered in some of the games I posted."
++ That proves those games are not relevant to weakly solving Chess.
"In ICCF games the players preder to agree a draw or play a triple repetition because they get to the pub quicker." ++ You have no clue. They play for years on a game.
This for example is the last finished game: ends with a perpetual check i.e. 3-fold repetition https://www.iccf.com/game?id=1164320
"Totally irrelevant to your calculations." ++ Totally relevant as example of perfect play.
"I don't propose a drawn KRPP vs KRP position in particular." ++ I do.
"If your calculations work they should work irrespective of material or theoretical result."
++ No. We are talking about weakly solving Chess, thus about optimal play by both players. Positions that do not result from optimal play are not relevant: they do not show up.
"you try your calculations on each of the sets of games I posted"
++ I posted a couple of those above and pointed out you have a problem with your Stockfish version. On my desktop the top 1 move coincides with a table base exact move.
"not just the games from the drawn KRPP vs KRP position I posted."
++ You did not post any drawn KRPP vs. KRP: the one you posted is a win.
"Black can draw from the initial position without the 50-moves rule, so this position cannot be reached from optimal play by both sides" ++ Yes, that is right.
"Optimal play by both sides I would take to mean no errors." ++ Yes, that is right.
"whether the calculations relate to basic rules errors or competition rules errors"
++ Errors are errors. The 50-moves rule is ignored as it plays no role.
Oct 12, 2022
0
#4618
That's now about a foot and a half of wriggling altogether. I'll respond later.
@6044
As you did not yet post any drawn KRPP vs. KRP, I post one:
The engine top 1 move is always table base exact.
This is not surprising: as most chess positions are at 26 men, it is only logical that most errors are around 26 men too.
The top 4 engine moves contain the table base exact move.
Whatever the number of games of chess, it's still a BIG number, comparable to 52 factorial.
Rather amusing point about the video is that it goes wrong almost immediately. First it says that if you have a room of 23 people, there is about 50% chance of two people having the same birthday. This is correct. Then it says, "so if you walk into a room with 23 people in it, the probability of someone having the same birthday as you is 50%".
No!
Specifying you as one of the two people leaves only 23 possible pairs that could match rather than the (23 * 22 / 2) pairs that the correct fact refers to. The chance of one of the 23 people having the same birthday as you is about 1 - (364/365)^23 ~= 6% (ignoring leap years).