Since you’re here...

We hope you will consider supporting us today. We need your support to continue to exist, because good entries are more and more work time. Every reader contribution, however big or small, is so valuable. Support "Chess Engines Diary" even a small amount– and it only takes a minute. Thank you.
============================== My email: jotes@go2.pl



New version chess Engine: Uralochka 3.41



Uralochka - UCI chess engine
Author: Ivan Maklyakov  Rating Chess Engines Diary CEDR=3616

UCI-compatible chess engine using Alpha-beta search and NNUE-like neural network.
The engine implements:
Basic control options (search to depth, time per move, time per game, time per N moves).
Support for Syzygy endgame tables.
Multithreading.
Builds for different processor extensions (SSE, AVX2, AVX512) and operating systems (Windows, Linux).

Engine architecture (everything is like most engines of this level):
Position representation and move generation - magic beatboards.
Sorting moves: moves from the transposition table, good captures by SEE, killer moves, quiet moves by history, bad captures by SEE.
Search - PVS, Aspiration search, Quiescence Search. And all possible abbreviations and cuts in the search (Null move, Static null move, abbreviations for quiet moves, Futility pruning, SEE prunung, LMR for quiet moves).
Transposition table for sorting moves, cuttings and for evaluation caching.
Multithreading - Lazy SMP.
Evaluation - a neural network similar to NNUE of the first versions with HalfKP. Network architecture: (2x(12288x1024))x6. Input layer: KingAreaPieceColorPieceSquare 166264. Trained using the frameworks Keras (early versions) and PyTorch (latest versions) on data generated in games of the engine against itself (with added positions from games against other engines) with ultra-short control (8-9 half-moves). Implementation of inference in the engine using the accumulator principle and vector instructions (SSE, AVX).
HCE (not used now) - material, PSQT, king's pawn shield, attacks on the king, pawn structure, piece mobility. The weights were selected using the Texel engine method.


v.3.41:
Changes:
Tuning the search function using SPSA
New network
Fixed time manager

Elo   | 140.92 +- 5.21 (95%)
Conf  | 10.0+0.10s Threads=1 Hash=32MB
Games | N: 10002 W: 4769 L: 921 D: 4312

Elo   | 69.03 +- 4.86 (95%)
Conf  | 60.0+0.60s Threads=1 Hash=64MB
Games | N: 10000 W: 3520 L: 1559 D: 4921


Uralochka 3.41 dev.3 - results: 
Ethereal 14.2520/44-444 Games
Rebel 16.320/44-444 Games
RubiChess 2024011219/44-644 Games
Chess-System-Tal-218/36+036 Games
Clover 6.1.1914.5/34-534 Games
Alexandria 6.1.011/30-830 Games
Seer 2.8.014/28+028 Games
Vafra 14.12.1 Stockfish13/24+224 Games
Velvet 7.3.011.5/24-124 Games
Berserk 1310/24-424 Games
Obsidian 12.010/24-424 Games
Viridithas 12.0.012/22+222 Games
RofChade 3.110.5/20+120 Games
Obsidian 11.09/20-220 Games
Caissa 1.178.5/20-320 Games
Lizard 10.310.5/18+318 Games
Black Marlin 9.010.5/18+318 Games
Caissa 1.185.5/18-718 Games
Velvet 7.0.07.5/14+114 Games
Koivisto 9.26.5/14-114 Games
Equisetum 1.08/12+412 Games

Comments