SugaR: Updated August, 13 2019.
After the latest surprising great performances (a tester of mine tried SugaR against Raubfisch, with a surprising result of Sugar doubling the wins), I couldn't avoid to host this engine on my website, but before I had to put my hand on it since I don't like all those files created in the same place where the chess engine is. So now they are all into a folder called 'SugaR-NN Files', a much more orderly way.
- Changelog: this distribution is an adaptation of SugaR-NN dated Jul 31, 2019, I have set the program to put any output file into the created directory 'SugaR-NN Files'.SugaR is a free UCI chess engine derived from Stockfish made by Marco Zerbinati under the GPL license.
This version of SugaR supports up to 128 cores; the engine can use two parallel BIN books (original code by Thomas Zipproth) and has a Self Learning function implemented.
NN section (Experimental Neural Networks inspired techniques)Experimental, MonteCarloTreeSearch, if activated, the engine's behaviour is similar to AlphaZero concepts.
Idea are implemented, integrated on SugaR:
NN Persisted Self-LearningBoolean, Default: True
(Montecarlo by Kelly Kinyama) only when true. This creates three files for machine learning purposes: SugaR-NN implements a persistent learning algorithm by Kelly kyniama and Andrea Manzo. Reads and creates the following file types: pawngame.bin with the learning when there are max a total of 2 pieces for white and black
experience.bin with the learning for
- opening variation of max 16 moves (8 half-moves) and a total of at least 7 pieces (no pawns) for white and black
- positions with max 6 pieces (no pawns) for white and black
One or many .bin files, each one associated to a single position biunivocally associated to the (technically, hashKey), in an opening variation of max 8 moves (16 half-moves) and a total of at least 7 pieces (no pawns) for white and black. This position is also in the experience.bin. So, these files are to speed the load in memory.
Every .bin file is so a collection of one or more positions stored with the following format (similar to in memory Stockfish Transposition Table):
- best move
- board signature (hash key)
- best move depth
- best move score
At the engine loading, there is an automatic merge to pawn.bin and experience.bin files, if we put the other ones, based on the following convention:
.bin
where
- fileType="experience"/"bin"
- qualityIndex , an integer, incrementally from 0 on based on the file's quality assigned by the user (0 best quality and so on)
The opening files can be simply copied and, in case of conflict/same name, the user must choice the one to use.
Because of disk access, to be effective, the learning must be made at no bullet time controls (less than 5 minutes/game).
NN Perceptron AlgorithmBoolean, Default: False
(Perceptron Sigmoid activation by Stefano Cardanobile) for Late Move Reductions search as training signal
NN MCTS ScoreBoolean, Default: False
((Montecarlo Tree Search Scores) by Jörg Oster) in main search function to an upper node.
This edition includes:
Windows:
- SugaR-NN_2019-08-13_32bit_general.exe for general 32-bit CPUs
- SugaR-NN_2019-08-13_32bit_old.exe for old computers
- SugaR-NN_2019-08-13_32bit_ppc-32.exe for 32bit power pc
- SugaR-NN_2019-08-13_32bit.exe for standard 32bit CPUs
- SugaR-NN_2019-08-13_x64_bmi2.exe for Haswell CPUs
- SugaR-NN_2019-08-13_x64_general.exe for general 64bit CPUs
- SugaR-NN_2019-08-13_x64_modern.exe for modern computers
- SugaR-NN_2019-08-13_x64_ppc-64.exe for 64bit power pc
- SugaR-NN_2019-08-13_x64.exe for standard 64-bit CPUs