Chess will never be solved, here's why

#3284
Weakly solving chess is the ultimate chess analysis: white tries to win, black tries to draw. Whenever white fails to win, white tries something else.
When white has exhausted all reasonable attempts to win, chess is weakly solved.

Somebody like Sveshnikov is most qualified: experienced in analysis without engines, experienced in analysis with engines, author of opening analysis books, grandmaster, world champion 65+.

60 moves is not "exceptionally long" for one of my blitz games.

357 move game - Leela versus Stockfish 2020 

"3) Different move orders leading to the same position are different games irrespective of the final position, so again this point serves only to show it's own irrelevance"
Not only different games - they can spin off into other games and positions too.

Defining 'solving' chess well - would include defining the enormity of the task.
But without illegitimate 'cutdowns'.
In other words - the whole task is set out objectively.
With all its sub-tasks and sub-projects.
No pretenses.

Analogy - somebody at Lockheed-Martin wants to cut corners.
Boss:  "Hey skip 90% of those project proposals.  Now !"
Engineers/mathematicians/scientists:   "OK Boss !"
Then it comes time for the first test flight of the fighter jet.
Government officials there.  Competing executives from Grumman and other companies ...
Does the engine even start up ?
Does the plane get off the ground?
Lets say by some miracle it did ...
On the very first bank - the plane simply slides down into the ground.
"Hey it was a 'weak solve' "
I would want neither tygxc nor Sveshnikov in charge of developing the aircraft.
No 'skipping'.  No cutting corners.
By the way - the death of the pilot would lead to indictments against 'the Boss'.

If poorly justified stubbornness was the key to solving chess, @tygxc would have done it by now.

Here here !  And Hear Hear !
Jolly Good post and Good Show !
Really.  Most efficient !

@Optimissed re #3292

I intended to say @tygxc, not you, misquoted Tromp (quite apart from comparing your figure intended to be games generated with Tromp's figure of legal basic rules positions).

Tromp actually estimates the number of basic rules legal chess positions as (4.82 +- 0.03) * 10^44 not 10^44. @tygxc routinely throws away the majority of whatever figure he arrives at for positions of various sorts, this being a minor example. The problem is it's infectious; the same figure of 10^44 is appearing in other users' posts.

If @tygxc intends to solve a version of chess that includes the 50 move and triple repetition rules then the number of basic rules legal chess positions would be a minuscule fraction of the nodes in his game tree and Tromp's figure would probably have no relevance to his proposed solution.

If he really has a proposed solution. 

#3295
4.82*10^44 is the same as 10^44 for all practical purposes.
Up / down symmetry and left / right symmetry after loss of castling rights reduce a factor 4, so that leaves 1.21*10^44. For all practical purposes it is 10^44 legal positions. That would be the number of positions needed for strongly solving chess i.e. a 32-men table base. Time and storage are prohibitive with present technology.

For weakly solving chess that figure has no relevance, as the vast majority of the positions Tromp found legal cannot occur in a reasonable game with >50% accuracy and thus certainly not in an ideal game with optimal moves i.e. perfect play. Those positions have multiple underpromotions to pieces not previously captured from both sides. Each such underpromotion represents a blunder worth a piece. Moreover underpromotions from both sides are not consistent: game-theoretically it makes no sense that both sides avoid stalemate.

A better figure is the 10^37 as calculated by Gourion.
Even the vast majority of these positions is not sensible: they cannot occur in a game of > 50% accuracy. Even after adding some positions with 3 or 4 queens there only remain about 10^32 legal and sensible positions.

The majority of these positions cannot be reached during the solving process. E.g. after 1 e4 all positions with a pawn on e2 are unreachable. That leaves about 10^20 legal, sensible positions reachable during the solving process.

The majority of reachable positions is not relevant. If 1 e4 e5 draws, then it is not relevant if 1 e4 c5 draws as well or not. If good moves cannot win, then bad moves cannot win either. That is why 1 a4 or 1 e4 e5 2 Ba6 are not relevant. That leaves 10^17 legal, sensible, reachable, and relevant positions.

3 cloud engines of 10^9 nodes/s can thus weakly solve chess in 5 years.
An alternative would be a cluster of 3,000 desktops of 10^6 nodes/s.
Another alternative would be Stockfish translated from C++ to Python and running on a quantum computer rented from D-Wave, Google, or IBM.

#3298

"you should have said FOR BASIC RULES CHESS. This is not chess as it is played."
++ As said many times: the 3-fold repetition rule is essential and is invoked in 16% of ICCF WC draws. It can be simplified to a 2-fold repetition rule for the purpose of solving chess. The 50-moves rule is never invoked in positions > 7 men and thus can be considered unwritten for the purpose of solving chess. As said many times: a position (FEN) is a diagram with the side to move, castling and en passant flags. To verify if two positions are the same for the 3-fold repetition rule, it is checked if the new FEN has previously occured twice. See Laws of Chee 9.2.2

"Firstly, this is a guess. Secondly it is a guess based on an undefined concept - accuracy."
++ No, it is neither a guess, nor an undefined concept. None of the 56011 sampled positions Tromp found legal can occur in a reasonable game because of the multiple underpromotions to pieces not previously captured they contain, even on both sides. Each underpromotion represents a loss worth a piece. That is clear at first glance of the sampled positions, but to make it objective load the pgn of the proof game that proves legality into the analysis tool in this site and read the accuracy figure. The accuracy figure does not prove that all moves are optimal, but if the accuracy figure is < 50%, then that proves that the moves are not optimal.

#3301
That is not a guess. Look through ICCF games, look at any data base.
Solving chess is chess analysis.
If the 'chess solver' invents nonexistent obstacles, then he will get nowhere.

Adding the word "formally" does not change things: it's unproven. Period. You propose to use the conjecture to reduce the search space to 10¹⁷ nodes. The evidence is not extraordinary enough to support such extraordinary claim. But don't make me repeat that. It's only the first problem.

That's not true, and it is proven: Lc0 at 1 node plays already at master level and can beat many engines rated less than 2100; back in 1995 Fritz 3 running on a Pentium at 90Hz won the computer championship ahead of Star Socrates and Deep Blue, the latter using 14 processors
capable to calculate 5 million positions per second.

Different moves may be equally effective, but it is well known that the stronger a player, the lower (on average) their ACL (average centipawn loss) to the best available engine. Therefore a game between to equally strong players resembles more and more a game autoplayed by the best engine, the stronger they are. Hence their systematic errors cannot be uncorrelated.

As for "thick" and "thin" nodes, the evaluation of a position is for the major part based on experience. This experience can be acquired both from past complete (but always imperfect) games played from similar positions, or from incomplete games played "on the fly", from the current position, in the player's mind (through calculations). The effects are comparable: the greater the experience, the "better" (more effective) the evaluation, but also more biased.

The information available is the knowledge, which makes the player choose A in that example. New informations, might alter the expected outcome of A.

Data are factual informations, like the increasing percentage of draws in ICCF games. Your error rate cannot be directly measured, so it cannot be considered a datum. You are as usual very careless about these distincions, if not blatantly unscrupulous. Plus, I did not say I cannot tell which is the error rate, I said it's impossible to tell, because they are not statistically independent. What's your estimation, if errors are not statistically independent?

Please, @tygxc: decades passed talking about the accuracy of engines, and no one has found 5 minutes do your calculations and derive your error rate per move?

SF on modern desktops already reaches 10⁷ nodes/second, so why do you insist so much on cloud engines is beyond me (or maybe not). But that's the least of the problems, anyway.

#3303

"in 1995 Fritz 3 running on a Pentium at 90Hz won the computer championship ahead of Star Socrates and Deep Blue, the latter using 14 processors"
++ 1995 is 27 years ago. Stockfish defeated LC0.

"Different moves may be equally effective"
++ Different human or engine players have different strengths and weaknesses.
An oversight by one does not correlate to an oversight by the other.

"The information available is the knowledge, which makes the player choose A in that example."
++ The expected outcome of hanging a piece is a loss.
However sacrifices can be correct and yield an expected outcome of a win.

"I said it's impossible to tell" ++ I say it is possible to tell: 126, 9, 1.

"because they are not statistically independent"
++ I say they are statistically independent based on the available games. Assume they are statistically independent, plausible as they are two different entities ICCF grandmasters with engines. Now calculate the result: 126, 9, 1. Now note there are only 9 single errors, independent while alone and one double error, independent or not. Thus the initial assumption is valid in this case.

"no one has found 5 minutes do your calculations and derive your error rate per move?"
++ So it must be wrong because nobody else has thought of it before?

"SF on modern desktops already reaches 10⁷ nodes/second"
++ Then the alternative would be a cluster of 300 modern desktops.

here we go again with another trash forum

And previously A0 and Lc0 defeated SF. In hystory there have been several examples of "thicker" vs. "thinner"nodes, sometimes the first approach prevailed, sometimes the other.

I have given a detailed explanation. You don't even object and just claim otherwise.

That's not an objection too and does not add anything to what I said.

Your thinking is circular, I do not know how to tell you. The example you made before:

is different, because the classical model is validated by experiments, for v << c, independently of the initial assumption (in fact it was thought correct for any v, before Einstein). Your reasoning is closed in itself: the result confirms the assumption with no external source of validation. Let's say coherent, but fallacious. Let's make another example: "I assume that people never meet each other in the room R. Therefore, no more than one person can be in R at any given time. Since no more than one person can be in R at any given time, people never meet each other in R".

I am not saying that. Since you are so fond with probabilities, I am asking you if you think it is plausible that, assuming your calculations correct, nobody else has thought of it before,

So Schaeffer solved checkers as an hobby (you say that!) starting the project with 200 computers and letting them run for 20 years, but he (or some other hobbyst) do not find the resources to start a project to solve chess in 20 years using 300 computers. Do you think it's plausible?

#3307

"And previously A0 and Lc0 defeated SF"
++ It was a crippled version of Stockfish. TCEC is good version against good version.

"sometimes the first approach prevailed, sometimes the other."
++ No, with short time control thick nodes are better. For solving chess the aim is to hit the 7-men endgame table base and thus thin nodes are better for that purpose.

"Different moves may be equally effective"
++ You give an erroneous explanation: the humans play more like engines and thus it is like autoplay. No, humans over the board and ICCF have different strengths, different weaknesses and play differently. Human vs. human is never like engine autoplay. In ICCF even less: they use different engines and the human decides. In Carlsen - Nepo Nepo made 4 clear errors, Carlsen jumped at it and won 4 times. I can give examples from ICCF: won games where one side clearly erred and the other side jumped at it.

"Your thinking is circular"
++ Yes, in a way it is, but it is usual in many sciences. Assume something a priori.
Calculate using that assumption. Verify the assumption is valid a posteriori . Calculation OK.

"because the classical model is validated by experiments, for v << c"
++ I deduce the error distribution from the plausible assumption of statistical independance. Then I calculate the error distribution. Then I verify that the error rate is low enough to validate statistical independence. That validates the assumption and the distribution.

"if you think it is plausible that nobody else has thought of it before"
++ Yes, it is plausible. Maybe somebody else thought of it before but did not care to communicate. Maybe somebody else communicated it somewhere before but we did not read it. Maybe nobody thought of the possibility. By your logic no patent would ever be awarded: either it is not new, or it is new and thus not plausible as nobody has thought of it before.

"So Schaeffer solved checkers as an hobby (you say that!) starting the project with 200 computers and letting them run for 20 years" ++ His main effort was generating his 10-men endgame table base. Later he had a reduced number of engines.

"he (or some other hobbyst) do not find the resources to start a project to solve chess in 20 years using 300 computers." ++ Schaeffer was not into chess. Maybe somebody does it with a cluster of 300 desktops of 10^7 nodes/s. It is a major hindrance to coordinate 300 desktops, but it is doable. The main difference between Chess and Checkers or Losing Chess is that Chess requires more knowledge. Also Losing Chess made use of knowledge. As the 7-men endgame table base is already available, weakly solving chess is a chess analysis problem.

"even I"
What happened there?
That looks like a blunder.   Introspection ??   Whaat??  
Major concession?
Kind of like one of tygxc's rare concessions ...
he made a major one at one point - during those days when his repetitions of invalid were not quite as nauseous ...
I and btickler both caught it.  Just can't remember what it was.

Not necessarily, because "thin" nodes mean less selectivity, and thus the engine spend more time on less promising lines.

As I said, the fact that they play differently does not mean they are not equally biased.

No.

No, because your verification process is just like the one in the example of the room.

Maybe.

They are usually a little more difficult to conceive than your deductions. As you say, most people think that the game-theoretic value is a draw and it is common to assume statistical independence to simplify calculations. Yet nobody did it...

But no one started such a project...

Sure they're the same repetitions.  Just not as nauseous.
But the point is the concession he made.
Perhaps a unique moment.
His sole moment of direct admission after several months and 3300 posts.