Page 21 of 28

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 11:39 am
by AndrewGrant
George Tsavdaris wrote:
Mon Sep 17, 2018 11:37 am
Not really.
I heard this second hand, so it seems I am wrong.

Maybe there is an even newer Fizbo, I don't know.

Sorry for the inaccuracy.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 1:44 pm
by jp
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 2:23 pm
by mjlef
jp wrote:
Mon Sep 17, 2018 1:44 pm
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.
I know Alpha Zero adjudicated the games it played versus Stockfish. Looking at a few of the 10 games they published, they do seem like losses for Stockfish. But I if Alpha Zero would have been able to finish the game. If you play until ahead some amount, will a nn engine be able to play to the end accurately all the time? I suppose this could be tested, but maybe this weakness might have not been found without playing to the actual end. Perhaps nn engine should be trained to the end of the game, or until they see mate or at least a draw score for a while.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 2:33 pm
by jkiliani
mjlef wrote:
Mon Sep 17, 2018 2:23 pm
jp wrote:
Mon Sep 17, 2018 1:44 pm
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.
I know Alpha Zero adjudicated the games it played versus Stockfish. Looking at a few of the 10 games they published, they do seem like losses for Stockfish. But I if Alpha Zero would have been able to finish the game. If you play until ahead some amount, will a nn engine be able to play to the end accurately all the time? I suppose this could be tested, but maybe this weakness might have not been found without playing to the actual end. Perhaps nn engine should be trained to the end of the game, or until they see mate or at least a draw score for a while.
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 4:10 pm
by Milos
jkiliani wrote:
Mon Sep 17, 2018 2:33 pm
mjlef wrote:
Mon Sep 17, 2018 2:23 pm
jp wrote:
Mon Sep 17, 2018 1:44 pm
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.
I know Alpha Zero adjudicated the games it played versus Stockfish. Looking at a few of the 10 games they published, they do seem like losses for Stockfish. But I if Alpha Zero would have been able to finish the game. If you play until ahead some amount, will a nn engine be able to play to the end accurately all the time? I suppose this could be tested, but maybe this weakness might have not been found without playing to the actual end. Perhaps nn engine should be trained to the end of the game, or until they see mate or at least a draw score for a while.
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.
And you still keep "0" in the name, what a joke...

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 4:28 pm
by CMCanavessi
Milos wrote:
Mon Sep 17, 2018 4:10 pm
jkiliani wrote:
Mon Sep 17, 2018 2:33 pm
mjlef wrote:
Mon Sep 17, 2018 2:23 pm
jp wrote:
Mon Sep 17, 2018 1:44 pm
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.
I know Alpha Zero adjudicated the games it played versus Stockfish. Looking at a few of the 10 games they published, they do seem like losses for Stockfish. But I if Alpha Zero would have been able to finish the game. If you play until ahead some amount, will a nn engine be able to play to the end accurately all the time? I suppose this could be tested, but maybe this weakness might have not been found without playing to the actual end. Perhaps nn engine should be trained to the end of the game, or until they see mate or at least a draw score for a while.
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.
And you still keep "0" in the name, what a joke...
I know you're just trolling, but (even if I don't like) TBs don't break the zero rule. It's not "human knowledge", it's perfect chess.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 4:30 pm
by Milos
CMCanavessi wrote:
Mon Sep 17, 2018 4:28 pm
Milos wrote:
Mon Sep 17, 2018 4:10 pm
jkiliani wrote:
Mon Sep 17, 2018 2:33 pm
mjlef wrote:
Mon Sep 17, 2018 2:23 pm
jp wrote:
Mon Sep 17, 2018 1:44 pm
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.
I know Alpha Zero adjudicated the games it played versus Stockfish. Looking at a few of the 10 games they published, they do seem like losses for Stockfish. But I if Alpha Zero would have been able to finish the game. If you play until ahead some amount, will a nn engine be able to play to the end accurately all the time? I suppose this could be tested, but maybe this weakness might have not been found without playing to the actual end. Perhaps nn engine should be trained to the end of the game, or until they see mate or at least a draw score for a while.
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.
And you still keep "0" in the name, what a joke...
I know you're just trolling, but (even if I don't like) TBs don't break the zero rule. It's not "human knowledge", it's perfect chess.
So perfect chess knowledge is equal to zero knowledge, lol. Your logic is really irrefutable :lol: :lol:.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 5:46 pm
by MikeGL
jp wrote:
Mon Sep 17, 2018 1:44 pm
Gary Internet wrote:
Mon Sep 17, 2018 6:19 am
They need to implement tablebase adjudication for this tournament to cut down o the number of times we end up watching engines messing about for 100+ moves of pointless crap.
At least 6 man if not 7 man adjudication
The problem is Lc0 is really bad at endgames.
But I think Lc0 with TB or syzygy support would solve this problem easily.
Not sure if Lc0 at CCCC have TB support though.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 5:55 pm
by Jesse Gersenson
MikeGL wrote:
Mon Sep 17, 2018 5:46 pm
But I think Lc0 with TB or syzygy support would solve this problem easily.
Not sure if Lc0 at CCCC have TB support though.
No TB's for Lc0; the Lc0 team asked to have TB disabled for their engine.

Re: Chess.com 2018 computer chess championship

Posted: Mon Sep 17, 2018 6:15 pm
by chessdev
We may consider doing TB adjudication in the future. That said, our time controls are short, and I believe we've seen some interesting games and positions because we haven't adjudicated. I'm open to suggestions! Perhaps for TB draws... but for wins, let's see it played out!?