jp wrote: ↑Mon Mar 02, 2020 12:47 pm
It certainly looks like SFdev at that TC was playing fairly randomly
That's unassisted engines for you!
Of course, as a centaur, you would do a lot better. You'd enable tablebases and "win".
More importantly, as a centaur without tablebases, you would do no better than the unassisted engine.
Based on what we've seen so far, as TB I think we could give the TBless centaur 200 moves and he still would not win. And note that all White has to do to keep going is avoid losing or exchanging a Bishop immediately, which even a computerless human can do. You certainly are not forced to find a single winning move at any time (without 50-move worries).
jp wrote: ↑Tue Mar 03, 2020 3:16 am
More importantly, as a centaur without tablebases, you would do no better than the unassisted engine.
Really? As a centaur I can go to lichess, paste a fen, and get the TB data of any position. The cursed win that Uri posted? Can play all the winning moves, no problem. Any other position that would require tablebases? I just go there and check what are the TB moves. On the fly, without downloading or storing anything.
This just proves you have no idea what being a centaur is all about, at some point we don't even need to use an engine!
This is getting on a huge sidetrack, which I don't want, but why not? Some people might like to play endgames and not just look them up.
And of course it's reasonable to ban engines in tournaments for people who want that. The only reason people think CC is "dead" is because of engines.
White in that position, with or without humans, with or without computers, but without tablebases, will surely fail to win. But let's see if SFdev has gotten anywhere since last we saw it.
Ovyron wrote: ↑Mon Mar 02, 2020 10:50 pm
why would someone pick the one that needs precise moves from the winning side is beyond me.
This has been answered many times. It's because there may only be one winning way. You cannot just pick when and where your opponent will make a mistake, shifting the presumed opening theoretical draw to possibly a win that needs precise moves.
And it's not clear that "needs precise moves" is a good description of this example. White doesn't exactly need to be precise if I say as TB I'll give him 200 moves.
jp wrote: ↑Tue Mar 03, 2020 3:16 am
More importantly, as a centaur without tablebases, you would do no better than the unassisted engine.
Really? As a centaur I can go to lichess, paste a fen, and get the TB data of any position. The cursed win that Uri posted? Can play all the winning moves, no problem. Any other position that would require tablebases? I just go there and check what are the TB moves. On the fly, without downloading or storing anything.
You're completely missing his point. Maybe you want to miss it.
He gave a tablebase position as an example precisely because we know the theoretical value of the position. Looking at unassisted engine performance and at assisted engine performance (your technique to build the tree of nodes to search and your "tagging", but no TB), we can compare then to known perfect play and see if they throw the win by failing to take the knight within 50 moves. "But I can check the TB" is completely irrelevant.
Then, we can extrapolate that if there are 5/6/7-men positions where a human-assisted engine would still fail to deliver a correct sequence of winning moves, there are even more 8-men+ positions where the same is true.
And "I would find an easier win" doesn't work in positions where there is no win easy enough for your method.
The best centaur in the world with the best hardware would still fail to spot many winning positions as such, much less convert them.
Alayan wrote: ↑Tue Mar 03, 2020 5:06 am
"But I can check the TB" is completely irrelevant.
It's not, because... I can check the TB.
A position needs to be posted where TBs can't be checked at all, because it's now "your side" who assumes there's a middle game position where my methods wouldn't suffice (so I fail to find the win), where presumably faster hardware would suffice, but examples where I can check the TB don't extrapolate because being able to check TBs is part of "magic tagging" (you see promising lines and you avoid those that reach endgames).
The point of the method is being able to take a shortcut to what unassisted engine would show at high depth, and perform better than it for the cases where it likes a plan that doesn't work because of misevaluation while one can have more accurate evaluation by being able to check TBs online.
In other news, I've been unable to backsolve a mate in 32 to move 40 or before in Zenmastur's challenge, because the number of lines is just absurd, so I'm going to give up now.
(still waiting for mmt's move, in case he missed my last move)
zullil wrote: ↑Mon Mar 02, 2020 11:14 am
I doubt that Stockfish would fail to win from this position in a game. Here's one quick test of default Stockfish-dev self-playing. Haven't looked to see where suboptimal moves occurred.
Thanks, but maybe that's just clueless play as both attacker and defender and it's attacking self just then stumbles into mate.
Can you see whether SFdev makes any headway just analysing from the start position?
Already in progress (for too many hours).
Finally something has happened that brings a glimmer of hope ...
I've terminated the search after 6332175484277 nodes and 36 hours. Stockfish-dev has failed to demonstrate a solution. The interested reader is welcome to analyze Stockfish's final PV for correctness ...