#30
"In fact, the proportion of white to black wins grows NON-LINEARLY with playing strength; for example at an Elo of 3000 the ratio is about 2:1, but at 3400 it increases to about 6:1." Where did you get that?
AlphaZero, LC0 and Stockfish NNUE do not apply minimax, but monte-carlo search instead, so they could find a hidden way to win if there were one.
Engines alone are no proof, but ICCF correspondence of humans + engines covers for the weakness of the engines.
Also human classical world championship as prepared with humans + engines provides further evidence.
If Carlsen and his team believed that there were a win for white, then he would not have alternated between 1 e4, 1 d4, 1 c4.
Now you are demonstrating your lack of understanding of search trees. Monte Carlo is derived from (and ultimately inferior to, were it complete) the minimax algorithm. Minimax must reach an end state to be complete, and a complete minimax tree search will always produce perfect play. In a game like chess however, minimax is INCOMPLETE and abbreviated by substituting a static evaluation function for end states at the horizon limit, since it cannot possibly reach the end state for all branches within the lifespan of the universe on current hardware. That is where it becomes weak. Monte Carlo on the other hand substitutes "expected outcomes" at the branch terminals by running several utterly random playouts to the end and applying statistical methods to come up with an outcome expectation. But the selection of the final candidate move arises from the same min-max analysis of those expected outcome values.
Random playouts are not going to find "hidden wins."
Hey! You are just dumb at all ; ZugAddict In his profile, it is a link of a game 1 + 0 with stockfish level 8 ( 3000 ), they have 0 mistakes, 0 blunders, etc. Look it here : https://lichess.org/hSs2Zhfs
It's because Nb6 is a blunder, allowing the thematic sacrifice. Otherwise it should be =.
Looking at that game, 7. ... Nc6 is the first novelty. That’s Fairy Stockfish (FS) 11.2’s favorite move at depth 24. We then transpose back in the a French Defense line. 10. ... b5 is the first true novelty, and it’s a move FS 11 doesn’t consider once it‘s 20 ply deep (i.e. once it’s looking 10 moves ahead); it thinks 10. ... f6 is far better. Indeed, b5 isn’t even in FS 11’s top 4 list (it’s #3 with Stockfish 13, after f6 and Qb6).
That in mind, 10 ... b5 is probably the first real error here.
12. ... Nb6, which Dario19503 correctly points out is an error, is not in Stockfish 13’s top 3 list nor in FS 11’s top 4 list (at depth 23; i.e. 11 1/2 moves ahead). Stockfish must had been under time pressure when it made that error.
#35
It is bullet...
"In other sciences, empirical data alone is NOT proof."
Depends on the science. With biology, knowledge by and large comes from fuzzy heuristics and empirical data. With physics, we do not consider a theory true unless we have empirical data backing up our theory (which is why string theory is questionable: Even though the math makes sense and works, we haven’t empirically observed magnetic monopoles)
Even with math, we use empirical data to claim something is probably true when we don’t have a formal proof: We believe there are an infinite number of twin primes, but this hasn’t been formally proven yet. Until 25 years ago, we empirically observed that Fermat’s last theorem looked true until we finally came up with a formal proof.
@h2b2 I think it will draw
maybe.................
@Goldrider Thus the qualifier "alone." And I'm mainly talking about the hard sciences here -- physics, chemistry, to a lesser extent biology and geology -- because those are more apt comparisons to the solution of chess, a game of perfect information with a finite if massive number of discrete states.
Most data collected in science is either observational, forming the basis to form a hypothesis, or experimental, collected under controlled conditions designed to support or falsify a hypothesis. Experimental methods are designed to the extent possible to control for other potential explanations; they test the hypothesis' unique predictions. When that is not possible, the experiments are designed instead based on theoretical explanations of observation, and then attempt to reproduce the observation.
The state of computer chess right now with regard to the question of the solution to the game is observational, because the limitations of current technology preclude a direct test of the hypothesis.
“The state of computer chess right now with regard to the question of the solution to the game is observational, because the limitations of current technology preclude a direct test of the hypothesis”
It’s not a question of the limitations of the current technology. It’s simply is not computationally feasible to model all possible Chess games in our universe:
• There are 10 ^ 120 possible Chess games [1]
• There are 10 ^ 82 atoms in the observable universe [2]
To model Chess, we need to somehow make every particle in the observed universe store around 10 ^ 38 (a very huge number: 100 Trillion Trillion Trillion) Chess positions.
So, any conjectures about chess will need to be based on empirical observations.
[1] https://www.chess.com/blog/Pau/how-many-possible-games-are-there
[2] https://physics.stackexchange.com/questions/386966/how-many-particles-are-there-in-the-entire-universe#387182
This topic has been archived and can no longer be replied to.