Zigqueen 6.6.0 - new version chess engine
Zigqueen is a UCI chess engine written in Zig (0.15.2) with a from-scratch NNUE evaluation and a single-threaded alpha-beta search. Engine code is a clean-room implementation (see CLEAN_ROOM_RULES.md); the network is trained from the publicly published Stockfish NNUE training datasets.
Author: Stierms Rating Chess Engines Diary CEDR=3821
What's new?
Strength
3738 (95%: 3734–3742) at one thread, from 3,256 games against 22 engines
centred on 3700 on the CCRL Blitz list; our own estimate, not a CCRL rating.
6.5.0 scored 3717 against the same engines with the same openings: +21 (95%:
+15 to +26), game for game.
What's new
A more compact hash table: more positions in the same Hash setting.
Deeper search: from depth 6 on the engine tests whether the hash move
stands alone, and searches it deeper where it does and shallower where it
does not. It reports about two plies more depth in the same time; depth
figures are not comparable with 6.5.0's.
Won games are converted more often: once its score has reached +3.00 in a
game, the engine presses for progress against the fifty-move counter. The
score shown can sink while that counter rises and returns after a pawn
move or a capture; a position analysed on its own is searched without it.
A mate found stays a mate from move to move.
Move ordering remembers the reply that refuted a move after captures and
promotions too, and orders the root moves by the search each one needed.
Where a position has been searched before, the pruning rules trust that
result over the static evaluation.
A draw inside the search scores +0.01 or −0.01, not 0.00; a draw that can
already be claimed and a tablebase draw stay 0.00. The engine repeats less
readily in a drawn position, so drawn games run longer (about one move in
seven more in our gauntlet) and more of them end by the fifty-move rule.
With movestogo, the last move before a time control is no longer
planned with the whole remaining clock.
Fixes
go movetime answers within the movetime less Move Overhead; 6.5.0
ignored the overhead and answered late about one move in eleven.
A stop within milliseconds of go gets a searched move.
With a draw score, a move that repeats an earlier position of the game is
searched before it is preferred; the opponent need not repeat.
A stalemate with pieces on the board is a draw, a mate in one behind an
empty capture list is seen, and close to the fifty-move draw a static
value or an old hash score is no longer trusted.
Comments
Post a Comment