'Chessbase wrote:
Vladimir Kramnik and Viswanathan Anand will play a special 4-game match in Dortmund starting on July 14, as castling will not be allowed. The event will be supported by world-leading AI company DeepMind. Online audiences will get to experience novel AlphaZero insights first hand in the post-match commentary of the no castling tournament.
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
Google Deepmind PR. Current Stockfish 14 is very likely to be stronger than AlphaZero at no-castling chess, and it is very easy for the Stpckfish net trainers to train a net specifically for no-vastling chess in a few days.
Since everyone has Stockfish 14, I am sure we will see how it compares. There is more than enough time (July 14) to train the net as you said.
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
Madeleine Birchfield wrote: ↑Wed Jul 07, 2021 4:52 pm
Google Deepmind PR. Current Stockfish 14 is very likely to be stronger than AlphaZero at no-castling chess, and it is very easy for the Stpckfish net trainers to train a net specifically for no-vastling chess in a few days.
Can't you just set up the starting position and stipulate that castling is disallowed (e.g. as you would if the king had already moved in another position)?
Human chess is partly about tactics and strategy, but mostly about memory
Yep, can just start with no castling rights fen, or could disable in engine for a little more performance/efficiency.
The whole custom net business just reeks of opening book (biases) though. Thus I would expect A0 to offer better evals/moves for the first ~10 moves, after that well SF is SF...
towforce wrote: ↑Wed Jul 07, 2021 6:00 pm
Can't you just set up the starting position and stipulate that castling is disallowed (e.g. as you would if the king had already moved in another position)?
You could. Alternatively, it could be reached from the vanilla chess starting position with 1. Nf3 Nf6 2. Nc3 Nc6 3. Rg1 Rg8 4. Rb1 Rb8 5. Ra1 Ra8 6. Rh1 Rh8 7.Nb1 Nb8 8.Ng1 Ng8.
Both TCEC and CCC have had competitions in the past where engines played no-castling chess.
(almost) absolutely no one cares about 'no castle' chess...except over the hill GM's who don't want to keep up with/can't remember the theory that helped make them GM's.
But...okay.
Cornfed wrote: ↑Wed Jul 07, 2021 8:20 pm
(almost) absolutely no one cares about 'no castle' chess...except over the hill GM's who don't want to keep up with/can't remember the theory that helped make them GM's.
But...okay.
Chess.com's CCC cared enough about no-castling chess that they did multiple events with the top engines on no-castling chess.
jhellis3 wrote: ↑Wed Jul 07, 2021 6:18 pm
The whole custom net business just reeks of opening book (biases) though. Thus I would expect A0 to offer better evals/moves for the first ~10 moves, after that well SF is SF...
Chess.com's CCC showed a few months ago that not having a custom net trained on no-castling chess positions results in about equal strength in no-castling chess between Stockfish with NNUE and Stockfish classic without NNUE, in contrast to Stockfish with NNUE being about 100 elo stronger than Stockfish classic in vanilla chess. An 100 elo difference cannot be solely due to opening theory.
More likely: king position is very important in Stockfish's half-king-all neural network architecture. In vanilla chess, the king ends up for the most part on g1/g8 or c1/c8 due to castling. On the other hand, without castling, the king has to manually travel to become safe, ending up on f1, f2, g2, or even staying on e1 many times. Training a net specifically on no-castling chess would help Stockfish with better king positioning during the opening and middlegame.
However I still think that current Stockfish, with or without NNUE, is stronger than AlphaZero at no-castling chess. Current Dragon is probably also stronger at no-castling chess than AlphaZero.
