And too invalid - people will understand - and reject and refute.
Chess will never be solved, here's why
Sort:
tygxc wrote:
"most of the times evaluations do not differ much, hence they are correlated"
++ When they do not differ, it makes no difference as there are no errors.
The errors occur when they differ: one is wrong.
😦 ??
tygxc wrote:
"which is what I mean with "evaluation"" ++ Fair enough. I call that 'provisional evaluation' or 'subjective evaluation' in contrast with 'objective evaluation' or 'absolute evaluation': draw / win / loss derived from either the 7-men endgame table base, or a 3-fold repetition [ . . . ]
How can there never be an error when two subjective evaluations do not differ? They can be wrong regardless they differ or not. Otherwise, they would be objective.
#3394
"How can there never be an error when two subjective evaluations do not differ?"
++ Look at the TCEC superfinals.
When the two subjective evaluations do not differ, the games have the same result.
#3395
The 7-men endgame table bases is objective. 3-fold repetition is objective. Some endgames like with opposite colored bishops are objectively drawn etc.
tygxc wrote:
"How can there never be an error when two subjective evaluations do not differ?"
++ Look at the TCEC superfinals.
When the two subjective evaluations do not differ, the games have the same result.
Meaning?
haiaku wrote:
They can be wrong regardless they differ or not. Otherwise, they would be objective.
Jun 10, 2022
0
#2766
tygxc wrote:
#3394
"How can there never be an error when two subjective evaluations do not differ?"
++ Look at the TCEC superfinals.
When the two subjective evaluations do not differ, the games have the same result.
I am getting a bit annoyed at the glibness of your nonsensical proclamations, so I will be blunt To those of us who think at all logically, it is trivially obvious that where there is some discrepancy between the evaluation of engines, it will ALWAYS be possible to find two that agree, even though a third totally disagrees!
😂😂😂😂😂😂😂😂😂😂😂
Idk why i am laughing 😂😂😂😂😂😂😂😂 i am getting so much laugh after seeing this forum 😂😂😂😂😂😂😂😂
#3398
"where there is some discrepancy between the evaluation of engines, it will ALWAYS be possible to find two that agree, even though a third totally disagrees"
++ The TCEC superfinals pits the 2 best engines against each other. The engines are totally different and thus independent. Whenever both agree in their provisional subjective evaluation, the objective final result of the game confirms that both are right. When they differ, one is wrong and the other one wins. Look at the TCEC superfinals: they have graphs of both evaluations and also graphs of the table base hits.
I have more reason to be annoyed at the nonsensical criticism but I refrain from bluntness.
There are few here who think at all logically. Most not even understand what is trivially obvious.
tygxc wrote:
The TCEC superfinals pits the 2 best engines against each other. The engines are totally different and thus independent.
Subjective. Engines share a great amount of their codes; they base their evaluations on similar "knowledge"; different types of machine learning can give similar results, as Elroch pointed out long ago. It's like saying that we are totally different from cats, while in fact we share 90% of our DNA with them.
Whenever both agree in their provisional subjective evaluation, the objective final result of the game confirms that both are right.
You jump to conclusions! Even if the game-theoretic value of the game was a draw, it's equivalent to saying that if you consistently draw with an equally strong opponent and you both agree on your evaluations, you both play perfectly!
When they differ, one is wrong and the other one wins.
Jump to conclusion! If player₁ thinks it's a draw and player₂ thinks "win for me", why it cannot be a win for player₁ instead, that both fail to recognize?
Jun 10, 2022
0
#2771
I would say tactics detections and observations of positions are objective (especially if its to be done well). Got to look.
but calculations comparing carefully prepared sequences of moves leading to and including final move selection ... is subjective.
Has to be. Some moves cannot be played because only one move at a time is allowed. 'There can only be One!'. Subjective.
#3403
"Engines share a great amount of their codes" ++ Yes, that is true
"they base their evaluations on similar "knowledge""
++ No, that is not true. AlphaZero bases its evaluations on knowledge it generated itself from no other input but the Laws of Chess by playing 700,000 games against itself. Its chess knowledge is thus logically derived from the Laws of Chess, i.e. axioms in your lingo.
"we are totally different from cats, while in fact we share 90% of our DNA with them."
++ We do not play chess like cats, or like apes, with which we share 99% of DNA.
Identical twins that share 100% of their DNA play chess differently.
Even two identical Stockfish engines, but with different settings play differently.
"If player₁ thinks it's a draw and player₂ thinks "win for me", why it cannot be a win for player₁ instead, that both fail to recognize?"
++ The TCEC superfinals consist of 100 games with 50 imposed openings, intended to be slightly unbalanced. The game-theoretic values of the imposed opening positions can be a draw, a white win, or a black win. They play a mini-match of 2 games with one of the 50 imposed openings. A priori there are several possibilities:
A) One engine wins both games: 2 - 0. That does not happen. It proves that the game-theoretic value of the imposed initial position impacts more than the difference in strength of calculation / evaluation between the engines.
B) One engine wins a game and draws the other: 1.5 - 0.5. That is the desired result of the imposed opening. It shows that the winning engine calculates/evaluates the imposed initial position and the resulting positions better than the other.
C) Each engine wins a game 1 - 1. That shows the imposed opening is too unbalanced.
D) Both games are draws 1 - 1. That shows the imposed opening is not unbalanced enough.
The most interesting occurence is B).
Now we can look at the evaluation graphs of both engines. In case B) we observe there is a difference in evaluation at some point. One side has it right and wins the 2-game minimatch, one side has it wrong and loses the 2-game minimatch. 1.5 - 0.5
For cases C) and D) we observe that the evaluation graphs of both engines agree. That means in case D) the imposed initial position was a draw and in case C) the imposed opening loses by force for one side.
There have been several TCEC superfinals: one per year, each 100 games.
There is a clear correlation between diverging evaluations and 1.5 - 0.5 results and between agreeing evaluations and 1 - 1 results, either win + loss or draw + draw.
Jun 10, 2022
0
#2773
'Looking' is objective. At a chess position. Lots to see!
But !!
You can only make one move at a time in chess !
Its Subjective !
There Can Be Only One !!
tygxc wrote:
"Engines share a great amount of their codes" ++ Yes, that is true
"they base their evaluations on similar "knowledge""
++ No, that is not true. AlphaZero bases its evaluations on knowledge it generated itself from no other input but the Laws of Chess by playing 700,000 games against itself. Its chess knowledge is thus logically derived from the Laws of Chess, i.e. axioms in your lingo.
And other neural network engines do the same, and we humans have learned from all the games played over centuries. You said a lot of times that A0 confirmes our knowledge, now you say the opposite?
"we are totally different from cats, while in fact we share 90% of our DNA with them."
++ We do not play chess like cats, or like apes, with which we share 99% of DNA.
What does that mean? Chess is just one of the differences between humans and cats, like vertical pupils, but both species have a nose, two eyes, two ears... we both eat, compete, reproduce, feel emotions... the similarities are much more than the differences. Likewise, the similarities between engines are more than their differences.
Identical twins that share 100% of their DNA play chess differently.
Even two identical Stockfish engines, but with different settings play differently.
How much differently? I use the decreasing average ACL, to get a measure of how evaluations get closer and closer on average, as the rating increases. Do the two SF you speak of have the same rating? If you don't want to use ACL, please post the distribution of the difference between their evaluations, so we can see how much they actually differ, or the probability that on average one engine plays a move which is different from the other engine.
"If player₁ thinks it's a draw and player₂ thinks "win for me", why it cannot be a win for player₁ instead, that both fail to recognize?"
++ The TCEC superfinals consist of 100 games with 50 imposed openings, intended to be slightly unbalanced. The game-theoretic values of the imposed opening positions can be a draw, a white win, or a black win. They play a mini-match of 2 games with one of the 50 imposed openings. A priori there are several possibilities:
A) One engine wins both games: 2 - 0. That does not happen. It proves that the game-theoretic value of the imposed initial position impacts more than the difference in strength of calculation / evaluation between the engines. [ . . . ]
Jump to conclusion, followed by many others.
Now we can look at the evaluation graphs of both engines. In case B) we observe there is a difference in evaluation at some point. One side has it right and wins the 2-game minimatch, one side has it wrong and loses the 2-game minimatch.
Jump to conclusion. What is this "those who win are right, those who lose are wrong"? To infer the possible causes from the effects one has to do much better than that. Besides, can you be more precise? Post the reference to those graphs and make a statistic, don't say just "look at". Is it possible to say that on average in case B) the evaluations differ more each other, compared to the other cases? For how many plies in each game? How many cases disprove the thesis (i.e. at some point they differ significantly, but then they agree again)? How many B) cases, over the total?
#3407
"You said a lot of times that A0 confirmes our knowledge, now you say the opposite?"
++ No, I say a lot of human and AlphaZero knowledge is absolute knowledge that logically follows from the Laws of Chess without any bias. Therefore that knowledge is allowed and beneficial to use in weakly solving chess with the brute force method. That absolute knowledge without any bias is in contrast to some subjective opinions: some people prefer 1 d4, some prefer 1 e4, some prefer 1 c4. However absolute knowledge, not subjective opinion, is that 1 a4 is certainly not better than 1 d4 or 1 e4.
"Likewise, the similarities between engines are more than their differences"
++ Formula 1 cars have more similarities than differences as well, but some drive faster.
"I use the decreasing average ACL, to get a measure of how evaluations get closer and closer on average, as the rating increases." ++ Average centipawn loss is a relative measure, as it is measured in comparison to some reference engine with reference settings. When you play a game with Stockfish running on a 10^9 nodes/s cloud engine and then evaluate it with Stockfish on a 10^6 nodes/s desktop then it will show a certain AcPL and an accuracy < 100%
"Do the two SF you speak of have the same rating?" ++ No. That is one part of TCEC preparation: tuning the settings for optimal results against the expected opposition.
They usually pitch several versions of their program against each other and then pick the most promising version for the competition.
"please post the distribution of the difference between their evaluations"
++ See TCEC it shows hundreds of evaluation graphs.
"What is this "those who win are right, those who lose are wrong"?"
++ The level of 3600 is that high, that a single error decides the game, just like in ICCF.
"Post the reference to those graphs" ++ Use Google: TCEC superfinals
"make a statistic" ++ You can do that
"on average in case B) the evaluations differ more each other" ++ Yes
"For how many plies in each game?" ++ Usually they are more or less the same, but at some ply (the error) they start diverging and they stay divergent until the end of the game.
"How many cases disprove the thesis" ++ I have not found any.
Jun 11, 2022
0
#2776
tygxc wrote:
#3398
"where there is some discrepancy between the evaluation of engines, it will ALWAYS be possible to find two that agree, even though a third totally disagrees"
Good that your post has some sense in it.
++ The TCEC superfinals pits the 2 best engines against each other.
At that time. As time has passed, those engines have continually been replaced by superior ones, new versions, new branches and occasionally entirely new candidates. Of course, you believe (for no reason) this process stopped last Wednesday.
The engines are totally different
Ridiculous. Some of the engines have some common code. All the best agree on the large majority of moves, even the most different. I am not actually sure if Leela agrees more or less with Stockfish than Stockfish with some selected different conventional engine, but it doesn't really matter. They are all different, and all similar.
and thus independent.
You write that like you have a clue what it means. That is clearly not so.
Whenever both agree in their provisional subjective evaluation, the objective final result of the game confirms that both are right.
Rational people burst into laughter at this point. Enough said.
When they differ, one is wrong and the other one wins.
It's brilliant how you m,ake such daft claims in an entirely serious manner.
Look at the TCEC superfinals: they have graphs of both evaluations
All the evaluations are wrong (the only correct evaluations are Win/Draw/Loss. In the case that one of the engines loses, the evaluation is wrong to the maximum possible extent .
and also graphs of the table base hits.
Table bases provide reliable evaluations (Win/Draw/Loss) of the final position. Only exhaustive analysis to the tablebase does as well.
I have more reason to be annoyed at the nonsensical criticism
No, you have more reason to overcome the limitations of your inflexible brain and to learn something.
but I refrain from bluntness.
There are few here who think at all logically. Most not even understand what is trivially obvious.
Some of us have spent large parts of our lives needing to be logical. You entirely fail to understand what is NOT trivially obvious, to a comical degree (see above).
In this post you have provided a rare example of a completely reliable proof. The firm conclusion is that you do not think in a rational manner.
Read my earlier post again and you would learn something if you could. It is unfortunate that has to remain purely hypothetical.
Jun 11, 2022
0
#2777
Use illogic constantly to get the attention of logic and to be corrected.
A kind of social life ?
Kids go through 12 years of pre-college schooling ...
with their homework and exams marked by teachers in red ...
adult attention ... almost like parent-figures.
Its now another 400 posts on the way to 4000 posts getting this kind of attention.
Extrapolating:
Say 4 months per 3000 posts.
So after 108,000 posts of this - that'll be 12 years of teacher attention.
Homework and exams - marked in red.
And we'll be in the year 2034.
Will the 5-year position ever pass a single exam - or even a homework?
In my school - homework was called a 'prep'.
tygxc wrote:
"You said a lot of times that A0 confirmes our knowledge, now you say the opposite?"
++ No, I say a lot of human and AlphaZero knowledge is absolute knowledge that logically follows from the Laws of Chess without any bias. Therefore that knowledge is allowed and beneficial to use in weakly solving chess with the brute force method. That absolute knowledge without any bias is in contrast to some subjective opinions: some people prefer 1 d4, some prefer 1 e4, some prefer 1 c4. However absolute knowledge, not subjective opinion, is that 1 a4 is certainly not better than 1 d4 or 1 e4.
It's impossible to argue with you, because you shift things all the time. If humans and A0 (and engines) share a lot of knowledge, how can they be totally different, as you say? Non sequitur.
"Likewise, the similarities between engines are more than their differences"
++ Formula 1 cars have more similarities than differences as well, but some drive faster.
How much faster? Are you saying that their speeds are uncorrelated? Because that's what we were talking about.
"I use the decreasing average ACL, to get a measure of how evaluations get closer and closer on average, as the rating increases." ++ Average centipawn loss is a relative measure, as it is measured in comparison to some reference engine with reference settings. When you play a game with Stockfish running on a 10^9 nodes/s cloud engine and then evaluate it with Stockfish on a 10^6 nodes/s desktop then it will show a certain AcPL and an accuracy < 100%
You just described ACL. The point is: if I find that the evaluations of two equally strong engines are correlated, how can their errors (i.e. they play a non-optimal move), which are based on those evaluations, be totally uncorrelated?
"Do the two SF you speak of have the same rating?" ++ No.
Then the greater the difference, the lower the correlation.
"please post the distribution of the difference between their evaluations"
++ See TCEC it shows hundreds of evaluation graphs.
You say they are not correlated! Should I do the homework for you? ACL is the result of statistics (average centipawn loss). Have someone made a statistic with the data you use, so we too can make use of it? Or you just looked at some of them? Can you post the link to the exact page/s you are talking about?
"What is this "those who win are right, those who lose are wrong"?"
++ The level of 3600 is that high, that a single error decides the game, just like in ICCF.
Circular thinking! So again the prove of statitstical independence between errors relies only on the assumption they are statistical independent. I am trying to determine if the statistical independence can be proven without assuming it as a premise, to have a source of validation outside the reasoning, otherwise it may be coherent, but does not prove anything, contrary to what you think. Only mathematical proofs need simply to be coherent, since they are exhaustive, and axioms are taken for granted.
Jun 11, 2022
0
#2779
Yes - he 'shifts' all the time.
But not much.
Only to the extent tactically necessary to continue with the Plan.
Jun 11, 2022
0
#2780
Optimissed wrote:
playerafar wrote:
Yes - he 'shifts' all the time.
But not much.
Only to the extent tactically necessary to continue with the Plan.
Which is to spread the 5-year virus?
Which is to create more 3000 post intervals each taking four months ... pushing the Snake Oil of 5 years and 'nodes'.
'Nodes' may as well be UFO's and 5 years may as well be 5 minutes in his 'treatise'
But even Gross illogic - so often has its own Internal Logic !!
#3388
"too concise"
++ Too long, people do not read.
Too short, people do not understand.
Some people do not read or understand either way.