Draw detection via Zobrist hashes in Position #11
Conversation
Test is currently breaking around a dozen plies in -- need to dig in further.
| @@ -167,6 +189,28 @@ instance Hashable Position where | |||
| `hashWithSalt` color | |||
| `hashWithSalt` flags | |||
|
|
|||
| epKey :: File -> Word64 | |||
…eplaced repetitions function
nvanderw
changed the title
| clockRemaining clock (color pos) >>= \case | ||
| Nothing -> pure (snd history, Win . opponent . color $ pos) | ||
| Just _ -> | ||
| if | draw pos -> pure (snd history, Draw) | ||
| | checkmate pos -> pure (snd history, Win . opponent . color $ pos) | ||
| | Just (n, _) <- repetitions poss | ||
| | n <- repetitions pos |
There was a problem hiding this comment.
This seems right given the new signature of repetitions but I'm unsure how to validate this code.
| @@ -1,42 +1,7 @@ | |||
| {-# LANGUAGE PatternSynonyms #-} | |||
There was a problem hiding this comment.
We discussed this a bit over email @mlang but to recap, circular dependencies prevent the Internal/Position logic from depending on Polyglot. Moving it here seemed simpler than duplicating the hashing logic, generating new hashes independent of Polyglot etc.
|
@nvanderw Glad to see you're working on this; ChessIO is a key library in a project I'm working on. Would you say your threefold-repetition detection is production ready? |
|
Hi. Sorry for the delay, I didnt spent much time on a computer during summeer time. I am about to merge this PR, however, during review I found |
Hey Mario, thanks for reviewing. Sorry for delay on my end, hope all is well for you. I think your read is right here and perhaps I was not awake yet when I wrote that piece. I think we can just use |
It is designed to be, and I've added test coverage to ensure the hashing works the way I expect, but I'm loath to claim that there are no bugs etc. I'd be curious to hear you try it out and tell me your experience. |
Currently chessIO has no efficient way to detect {three,five}fold repetition draws from a given position, i.e. to tell how many times the position has occurred. The way some modern chess implementations handle this is maintaining a position hash as part of the position data type, and also storing a linked list of hashes of the previous board states (see e.g. Stockfish). Since recomputing the hash of the entire board each move would be expensive, Zobrist hashing is used to quickly update the hashes with only the changed elements of the board.
(Optimization: in order to detect if a given position is a draw by repetition, we only need to know about the prior board states since the last pawn move or capture.)
chessIO already has some Zobrist hashing code in the Polyglot module, but no support for incrementally updating the hashes. So I've moved that hashing logic under Game.Chess.Internal.
After applying a move to a position in
unsafeDoPly, I've added a step to XOR the before and after bitboards together -- any 1s indicate changes to the board state. For each detected change to the board state,unsafeDoPlylooks up the corresponding hash in the Zobrist tables and XORs it with the position hash.As of 2022-4-12 I'm still ironing out some bugs in the hashing implementation. And then the next step is to implement the hash history updating/searching. But thought I'd send this out for an early review in the meantime. Also: exposing position hashes to library clients would allow them to use transposition tables.