Chess has been solved (by that I mean all endgames with up to seven pieces)
Chess will never be solved, here's why
Sort:
Mar 22, 2023
0
#8783
Kyobir wrote:
Chess has been solved (by that I mean all endgames with up to seven pieces)
Not quite.
No positions with castling rights even if you're talking about weakly solved. (And a lot of the Lomonosov DTM solutions are no longer freely available, if available at all, though that's not strictly relevant.)
MARattigan wrote:
We already have a viable method for shortcutting the process, it's called tablebase generation. Unfortunately, while it's a viable method (I should say viable methods) for shortcutting the process, it still doesn't result in a viable process.
For example a KNN v K tablebase contains around 12.5 million positions under FIDE basic rules (in most people's parlance), but less than a hundred will be included in a Nalimov tablebase (possibly a few thousand considered in total). By far the biggest reductions arise when the 50/75 move and triple/quintuple repetition rules are included.
Granted a factor of 4.82 is not particularly significant in terms of 10^44, but it should at least mean @tygxc starts talking about 11 years using 3 cloud computers + 7 maids with 7 mops instead of 5 years, if he's going to be consistent in his own terms. It would, of course, be an insignificant step in the right direction in real terms.
Tablebases are the solution, but not a shortcut, and we're already in agreement that they are the only viable path right now. Which is why chess will not be solved in our lifetimes.
Constructing a tablebase MEANS solving chess so it isn't so much a path as a framework to place the results. The only viable path is algorithmic and at the moment there is no prospect of developing an algorithm that can look at relatively simple positions (anything beyond the starting position is necessarily a simpler position than the starting position) and make a perfectly accurate prediction as to what would be the outcome of perfect analysis were it to be available.
To solve chess in any meaningful way we would have to be content with an accuracy projection which isn't 100%. Therefore how meaningful would it be?
Mar 23, 2023
0
#8787
DannyTheClauncher wrote:
What do you guys mean by solving chess?
It would be like the current 7-piece tablebase, only with 32 pieces from a particular starting point.
i agree
TheChessIntellectReturns wrote:
Imagine a chess position of X paradigms.
Now, a chess computer rated 3000 solves that position. All well and good.
Could another computer rated a zillion solve that position better than Rybka?
No, because not even chess computer zillion could solve the Ruy Lopez better than a sad FIDE master could.
the point is, there's chess positions with exact solutions. Either e4, or d4, or c4, etc.
nothing in the world can change that.
So if you are talking about chess as a competitive sport, then chess has already been solved by kasparov, heck, by capablanca.
If you are talking chess as a meaningless sequence of algorithms, where solving chess equates not to logical solutions of positional and tactical prowess, but as 'how many chess positions could ensure from this one?'' type of solutions, then, the solutions are infinite.
So can chess be solved? If it is as a competitive sport where one side must, win, then it has already been solved. Every possible BEST move in chess has been deduced long ago.
If chess is a meaningless set of moves, with no goal in sight, then sure, chess will never be solved.
True.
I'm born male and I identify myself as a man
Mar 23, 2023
0
#8792
DannyTheClauncher wrote:
What do you guys mean by solving chess?
You can spend a lot of time reading through this thread and you'll see plenty of discussion and disagreement about that. Some say you need only examine the "best" openings (e4, d4, c4, Nf3), use engine-aided correspondence games as "perfect" opening play, and have a committee of five GMs examine super-computer analysis and exclude lines they feel aren't worth thorough analysis to get a few chosen variations worked all the way to the seven-piece tablebase. Others maintain that only an analysis of every possible variation from the starting position to checkmate or a known draw will yield an irrefutable conclusion.
You can decide on your own criteria from wherever in that range you feel will be convincing.
Mar 23, 2023
0
#8793
DesperateKingWalk wrote:
ScarletDeathweaver2003 wrote:
I JUST BEAT STOCKFISH 15.1 earlier from the starting position WITHOUT ASSISTANCE OR STOCKFISH HANDICAP!! The fact that STOCKFISH 15.1 only plays perfectly (much better than any human being at least) makes Stockfish 15.1 more predictable so I prepared a line 58 moves (115 half moves) deep with all of the best moves because the last move in my preparation was a7 up 4 pawns for a bishop and there was no way to even stop that pawn from Queening on a8 so I was up a queen and 3 pawns for a bishop and no compensation for black!!
I do not know why you think you can falsify a game. And think you can get away with it.
The analysis clearly shows Stockfish 15.1 was forced to play suboptimal moves that Stockfish 15.1 would not play.
This is clearly seen in the centipawn analysis score for Stockfish 15.1 that scored a very weak .40 centipawn loss average per move. .40 is not even good for a human player, let alone the strongest chess engine in the world.
Then you have the nerve to claim your play was " WITHOUT ASSISTANCE OR STOCKFISH HANDICAP!!'
The analysis clearly shows you played with assistance. Having a centipawn score of .04, That is flawless play. And your move correlation was almost 100% with Stockfish.
And the analysis clearly shows Stockfish 15.1's play had almost no correlation with itself in your game. And the human player who is claiming the game was played without assistance has almost a 100% correlation with Stockfish 15.1.
That is not possible if the game is legitimate!
"centipawn"???
Mar 23, 2023
0
#8794
DesperateKingWalk wrote:
Kyobir wrote:
DesperateKingWalk wrote:
ScarletDeathweaver2003 wrote:
I JUST BEAT STOCKFISH 15.1 earlier from the starting position WITHOUT ASSISTANCE OR STOCKFISH HANDICAP!! The fact that STOCKFISH 15.1 only plays perfectly (much better than any human being at least) makes Stockfish 15.1 more predictable so I prepared a line 58 moves (115 half moves) deep with all of the best moves because the last move in my preparation was a7 up 4 pawns for a bishop and there was no way to even stop that pawn from Queening on a8 so I was up a queen and 3 pawns for a bishop and no compensation for black!!
I do not know why you think you can falsify a game. And think you can get away with it.
The analysis clearly shows Stockfish 15.1 was forced to play suboptimal moves that Stockfish 15.1 would not play.
This is clearly seen in the centipawn analysis score for Stockfish 15.1 that scored a very weak .40 centipawn loss average per move. .40 is not even good for a human player, let alone the strongest chess engine in the world.
Then you have the nerve to claim your play was " WITHOUT ASSISTANCE OR STOCKFISH HANDICAP!!'
The analysis clearly shows you played with assistance. Having a centipawn score of .04, That is flawless play. And your move correlation was almost 100% with Stockfish.
And the analysis clearly shows Stockfish 15.1's play had almost no correlation with itself in your game. And the human player who is claiming the game was played without assistance has almost a 100% correlation with Stockfish 15.1.
That is not possible if the game is legitimate!
"centipawn"???
Average centipawn loss is the number of hundredths of a pawn by which a player deviated from the most accurate move calculated by a computer.
so if you could have captured a queen but you left your queen vulnerable, it is -1800 centipawns?
Mar 24, 2023
0
#8795
It is incorrect to think of a centipawn loss as being an underlying truth. It is an evaluation by some specific engine with some specific time, indicating its lack of full understanding of the position. an engine that has full understanding could only give 3 evaluations - 0.0, +infinity and -infinity, depending on the theoretical result with perfect play (instead of infinity, engines actually state the moves to mate assuming the winner wants to minimise this and the defender wants to maximise it.
Where a weaker engine gives an evaluation, a stronger one may give a different one in either direction, by any amount.
Mar 24, 2023
0
#8796
From now on, let's refer to this "perfect chess engine" as Stockfish: Final Form. Now here are a few questions:
1) What would be this bot's ELO?
2) Would it only classify moves into Best Move and Blunder?
3) Would the creation of this bot end chess as we know it?
…
Kyobir wrote:
From now on, let's refer to this "perfect chess engine" as Stockfish: Final Form. Now here are a few questions:
1) What would be this bot's ELO?
2) Would it only classify moves into Best Move and Blunder?
3) Would the creation of this bot end chess as we know it?
No, let's not. That would confuse people who didn't know that "Stockfish" is being used as code for "Not-Stockfish".
Its ELO cannot be judged because that would depend on its style of play rather than its perfection. Work out the reason for that.
Mar 25, 2023
0
#8800
It is true that game theoretic optimality and Elo are to some extent independent. Indeed, I believe it is possible in principle for there to be a game theoretic optimal strategy that would not be difficult to draw against. But such a strategy would be very different to an extrapolation of Stockfish (or any other engines, including AIs). The reason is that these imperfect engines aim to make the position as difficult as possible for the opponent. This tends to mean they play optimally when winning as a side effect of the main objective (centipawns, or estimated expected score for an AI).
To generate an engine that is optimal but easy to draw against, you really need two things. The first is a strong solution (so you always know which moves are optimal). The second is a strong engine (by human standards) that is not perfect (so gives a wide range of evaluations in almost all practical drawing positions). The strategy used is then to always pick the LOWEST evaluation move that is game-theoretic optimal. This should make it fairly easy for the opponent to retain the draw. Roughly what is required is that the centipawn "errors" of the strategy are sufficiently large to balance the errors of the opponent. Of course centipawn errors are not fundamentally meaningful, but the idea is that if the evaluation does not rise too much about zero, the opponent will retail the draw,
For example, let's consider the first position. The optimal strategy with low Elo plays the lowest evaluation first move that does not lose. This may be 1. g4, but since the evaluation of this at depth 42, using Stockfish 14 is a miserable -1.7, it is possible that this move is a forced loss. If it is, our strategy will choose 1.Na3 or 1.h4, each of which have evaluation -0.7 with a very deep analysis.
Suppose 1. Na3 is played, then our average player feels moving a central pawn is likely to be ok, so picks 1. ...d4. According to Stockfish 14 this is the fifth best move, but still leads to -0.4 evaluation. In this position white has multiple moves with evaluations below -1.0 and there is surely a low evaluation, game theoretic optimal move that will make it easy for black to retain the draw.
And so on, with white continually trying to keep the evaluation as much below zero as possible without losing, until a comical draw if black avoids ever playing a losing blunder. The key is that in low evaluation positions there are enough plausible moves that are optimal for black to avoid playing a blunder.
======================================================================
Anyhow, this contrasts with the notion of the Elo that is the limit of the Elo of imperfect engines like Stockfish or LeelaZero as they approach perfection by good fortune as they attempt to maximise centipawns. This can even be defined in a way independent of any future software development - imagine a sequence of copies of some specific iteration of Stockfish that is provided with increasing computing power and memory (eg double them, over and over again). It is not unreasonable to believe this engine:
- approaches perfection
- its Elo keeps rising
- the Elo is bounded above
The last is a consequence of the finiteness of chess, I think. With a suitably constrained definition of player, you can't have an infinite sequence of players, each strictly superior to the previous one.
Not even if you're planning to solve chess according to current FIDE basic rules (no 50 move or triple repetition rules) which has generally been regarded as chess only since 2017.
Several points.
1. The best estimate of the number of legal positions under current FIDE basic rules is nearly five times the number you quote. I would guess the number you quote would not cover the legal positions pursuant to most KPP v KPP positions if the 50 move and triple repetition rules are included in the rules (depending on what you mean by "position").
2. It's not necessary to evaluate all legal positions to weakly solve chess. How many depends on the method of finding a solution.
3. First of all you need a relevant definition of "position" - relevant to some method of finding a solution, that is. It's easy to find situations in possible games under rules that include the 50 move and triple repetition rules where the FENs at the completion of a move are identical but the correct result in terms of win, draw or loss are different.
Ok, I'll put you down on the side of the guy who thinks 6 ply is enough to do the job .
Until a viable method gets put forth for shortcutting the process, the number is over 10^40. Whether that is 10^40, 10^42, 10^44, 10^46, etc, is immaterial...they are all impossible given current technology *and* any reasonably foreseeable technology. "5 times the number" is relatively insignificant at those orders of magnitude.
Cue the dweebs that think that quantum computers as they sit now can solve chess...
We already have a viable method for shortcutting the process, it's called tablebase generation. Unfortunately, while it's a viable method (I should say viable methods) for shortcutting the process, it still doesn't result in a viable process.
For example a KNN v K tablebase contains around 12.5 million positions under FIDE basic rules (in most people's parlance), but less than a hundred will be included in a Nalimov tablebase (possibly a few thousand considered in total). By far the biggest reductions arise when the 50/75 move and triple/quintuple repetition rules are included.
Granted a factor of 4.82 is not particularly significant in terms of 10^44, but it should at least mean @tygxc starts talking about 11 years using 3 cloud computers + 7 maids with 7 mops instead of 5 years, if he's going to be consistent in his own terms. It would, of course, be an insignificant step in the right direction in real terms.