Chess Engines Diary

Leorik - UCI chess engine Author-Thomas Jahn   Rating Chess Engines Diary CEDR=2998 Leorik 3.2 - what's new? Leorik 3.2 now supports Fisher Random chess. It also implements new UCI features like MultiPV and Pondering and gains strength from a more sophisticated NNUE architecture and a few search improvements. Built with .Net 10. The new NNUE architecture uses horizontal mirroring, 5 input buckets (king pos) and 8 output buckets (piece count). The network 640HL-S-5io8-6116M-FRCv1.nnue‎ supports both FRC and standard chess and was trained on over 6 billion labeled positions, roughly 1 billion of them from Fisher Random games. Further improvements include dynamic NMP reductions based on remaining depth. Improved time control that takes the stability of the search into account. And Razoring which uses adaptive margins based on live statistics. Leorik 3.0 vs other engines:   Booot 7.3 2.5/10 -5 10 games Alexandria 6.0 2/10 -6 10 games Leptir 070224 1/10 -8 10 games ShashChess 34.6...

Leorik - UCI chess engine Author-Thomas Jahn   Rating Chess Engines Diary CEDR=2998 v.3.1: Leorik 3.1 uses .Net 9 and has a stronger NNUE evaluation and an improved search. The NNUE improvements include: Larger network: Hidden-layer size increased from 256 to 640 neurons. SCReLU Activation: replaces CReLU for better accuracy. Trained from scratch: Training data collected from 19 generations of selfplay, starting with a zero-initialized network and high initial temperature. New Network: 640HL-S-5288M-Tmix-Q5-v19 trained on 5.2 billion positions, gathered across generation 13 through 19. The Search improvements include: Correction History: static eval adjusted by search results on positions with similar features. (Compares Pawn, Knight|Bishop and Queen|Rook bitmasks) Wiki Aggressive Late Move Reduction: Late quiet moves that do not improve the static eval are getting reduced by 4 instead of the usual 2. Dynamic Late-Move Threshold: The history threshold that identifies quiet moves a...

Leorik - UCI chess engine Author-Thomas Jahn   Rating Chess Engines Diary CEDR=2998 v.3.1: Leorik 3.1 uses .Net 9 and has a stronger NNUE evaluation and an improved search. The NNUE improvements include: Larger network: Hidden-layer size increased from 256 to 640 neurons. SCReLU Activation: replaces CReLU for better accuracy. Trained from scratch: Training data collected from 19 generations of selfplay, starting with a zero-initialized network and high initial temperature. New Network: 640HL-S-5288M-Tmix-Q5-v19 trained on 5.2 billion positions, gathered across generation 13 through 19. The Search improvements include: Correction History: static eval adjusted by search results on positions with similar features. (Compares Pawn, Knight|Bishop and Queen|Rook bitmasks) Wiki Aggressive Late Move Reduction: Late quiet moves that do not improve the static eval are getting reduced by 4 instead of the usual 2. Dynamic Late-Move Threshold: The history threshold that identifies quiet moves a...

Another group E-06 in the Amateur Chess Engine Section. The Leorik 3.0 engine won and the Pingu 4.0.0 engine came in second. These two engines advanced to the E-07 group. The last two engines in the table, Devel 4.0.2.3 and Protej 0.6.7, dropped to the E-08 group. More tests are underway, the results will be posted on our website soon. Chess Engines Diary Team 👍Promotion  to Group E-05:  Leorik 3.0, Pingu 4.0.0 . 👎Relegation to Group E-07:  Devel 4.0.2.3, Protej 0.6.7 .  All CEDR  515.173 games download  (01.12.2024)  💾  90 games from the tournament download 👍 @chessenginesdiary Country - Poland, City -  Malbork   🕓 Time 3'+3"  💻Asus Zenbook, AMD Ryzen 9 6900HX, RAM 32,0 GB 🖬 GUI-Banksia Tech table: Engine KN/move NPS dep/mov time/mov mov/gam time/gam fails 4ku 5.1 4594 962982 30.6 4.8 76.3 363.9 Ceibo 1.0 9349 1783850 30.4 5.2 62.9 329.6 1 Devel 4.0.2.3 3676 760014 22.8 4.8 78.5 379.6 2 Dumb 2.1 17...

Leorik - UCI chess engine Author-Thomas Jahn   Rating CEDR=2998 v.3.0: Leorik 3.0 replaces the previously used HCE with a NNUE based evaluation. The architecture is (768->256)x2->1 which means that the network uses 768 inputs (2 colors x 6 piece-types x 64 squares) and one hidden layer of 256 neurons. It maintains two separate accumulators (from black's perspective and from white's perspective) and has two sets of output weights to allow the network to learn tempo. It uses quantization (16bit integers) and AVX2 intrinsics so that 16 "synapses" can be processed in parallel. And of course the two accumulators are updated "efficiently", that is incrementally, by adding and subtracting weights when a piece moves or captures. The network was trained using Bullet on 622M labeled positions extracted from selfplay games. Rating CEDR progres - Leorik   Demolito 2023-08-20 2/16 -12 16 Games Caissa 1.13.1 0.5/16 -15 16 Games Carp 3.0.1 0.5/16 -15 16 Games Mida ...

Leorik - UCI chess engine Author-Thomas Jahn   Rating CEDR=2829 v.2.5 Leorik 2.5 includes a number of new features and improvements that boost it's playing strength by approximately 100 Elo. Here are some of the key updates: Open-Source: Leorik's source and binaries are now shared under the permissive MIT License. .Net 8 Migration: The engine has migrated from .Net 6 to .Net 8 for better performance and compatibility with the latest .Net features. Faster Move Generation: Leorik now employs a faster move generator based on PEXT by default but also includes the previous and other sliding-piece move generators, that are selectable via compiler define. Advanced Piece-Square Values: Leorik 2.5 replaces the standard PSQTs with linear functions that calculate the piece-square values using 18 parameters reflecting the positions of both kings and the game phase. Computing these formulas on the fly at acceptable speed requires AVX2 hardware support. AVX2 256bit wide registers allow the ...

Leorik - UCI chess engine Author-Thomas Jahn   Rating CEDR=2771 v.2.4 This version keeps the evaluation of version 2.3 but features an improved the search.  The most notable addition is the use of a Static-Exchange-Evaluation (SEE) to skip bad captures in Quiescence search and to search moves with a bad SEE score at a reduced depth in the main search. This new version also includes a nice simplification: I removed Futility Pruning and made Null-Move Pruning the only remaining place where entire branches can be pruned because they are just 'too good' for one side. Futility pruning relied on the static evaluation of a position and didn't explore it further if this evaluation was way below alpha. I now use the same idea in my null-move pruning implementation: When a position's static evaluation is way above beta and close to the leaves then Leorik just returns beta without further evaluation. Last but not least Leorik now recognizes certain drawn positions with a material...

Leorik - UCI chess engine Author-Thomas Jahn   Rating CEDR=2775 v.2.3 All previous versions had their evaluation tuned on a dataset of 725k annotated positions from Zurichess. The labels of these positions were derived by using Stockfish, so previously Leorik owed a part of his chess knowledge to two other engines. In this version all previously handcrafted evaluation terms have become tunable weights and all weights are tuned from scratch on selfplay games. The initial batch of selfplay games were played with a version that evaluated positions only using the commonly accepted Point Values of Chess Pieces. Even though this version lost about 700 Elo in strength the games were good enough to tune a new set of weights, compile a stronger version and repeat the process. After half a dozen iterations Leorik surpassed it's old playing strength. And it's showing a new aggressive playstyle discovered by playing thousands of matches against itself. In addition to this major change the...

Leorik - UCI chess engine Author-Thomas Jahn   Rating CEDR=2551 v.2.2 This version adds a mobility term to the evaluation: Each non-pawn piece receives bonus cp based on the number of non-capture moves it can make (up to a cap) multiplied with a small, constant cp value. The other changes address irregularities observed with the last version: Null-Move pruning in pawn endgames has been disabled because of the increased risk of missing Zugzwang. The replacement scheme of the Transposition Table has been rewritten to better protect deep nodes in matches using long time-control settings. The time-control logic has been completely rewritten to improve performance in matches without per-move increment. Leorik 2.2 gains about 100 Elo in strength. Individual statistics: Leorik 2.1 - 216 games CEDR (+ 59,= 13,-144), 30.3 %   Critter 1.6a                  :   2 (+  0,=  0,-  2),  0.0 % Houdini 1.5a  ...