Chess engine: Eubos 4.0 NNUE

Eubos - a basic Java chess engine, written for kicks. Uses jcpi for the UCI protocol parts. Kind of weak, doesn't evaluate for king safety, for example. Now uses transpostion hashing, but because it runs in the Java VM on a bog standard PC, doesn't search very deeply (around 100K Nodes/s, though this is actually quite hard to evaluate due to the hashing). 

Author: Chris Bolt Rating CEDR=2493

v.4.0:

This is the first release of Eubos as a Neural Net evalution chess engine.

Thanks to Dominik Klein for writing 'Neural Networks for Chess'. A really nice book that was the departure point for me writing Eubos v4.0. I love the spirit of this book and I was happy to buy the print version from Amazon. I can even forgive you the many typos ;)

I have to credit and offer thanks to Jamie whiting, author of the Bullet Neural Net training tool. His excellent code and documentation were invaluable. I must also thank and credit the author of Bagatur, from which I derived the basic inference code for accessing a Bullet trained neural net in Java. I have refactored that a bit to adapt it to my needs, but it really gave me a quick start. I also used the Bagatur v2 net whilst initially evaluating the approach. I also want to thank the author of Leorik, for his helpful forum posts on moving to a neural net eval and thoughtfully and gratiously sharing his experience in engine programming.

I have adapted the most primitive of the Bullet net architectures, for a 768 x 128 x 1 neural net. It is trained on 37 million positions, which were from a mixture of 1) Eubos self-play games to a fixed node depth and 2) randomly self-played games with each random selected move evaluated to a search depth of 8 plies. I stopped at 37 million in my training set because I can see the loss increasing and play strength stagnating. Therefore I deem this release sufficient for v4.0.

I have removed the v3.x hand-crafted evaluation. If I go back to this at any point it will be for a release from the 3.x codebase (which would putatively be v3.10, if I ever do it).

Eubos 3.9 JA vs other engines:

Pioneer 0.3.2 JA	4/4	+4	4 Games
Mufasa 0.2.1 JA	3.5/4	+3	4 Games
Integral 7.0.0	0/4	-4	4 Games
HypnoS ++ 1.02	0/4	-4	4 Games
GOOB 1.8.9 JA	0/4	-4	4 Games
Peacekeeper 3.01 JA	0/4	-4	4 Games
Odonata 1.0.0 JA	0/4	-4	4 Games
Obsidian 15.03 SE	0/4	-4	4 Games
Killfish PB 090225	0/4	-4	4 Games
Midnight 9 JA	0/4	-4	4 Games
Fatalii 0.9.0	0/4	-4	4 Games
Beef 0.3.6 JA	0/3	-3	3 Games
Bitbit 1.3	2/2	+2	2 Games
Rustic Alpha 3.05 JA	2/2	+2	2 Games
Reggz 0.6.0	2/2	+2	2 Games
Deep Blunder 1.2.0 JA	2/2	+2	2 Games
Jester 0.85 JA	2/2	+2	2 Games
Sloth 2.0 JA	1.5/2	+1	2 Games
100 Elo Chess Engine	1.5/2	+1	2 Games
Fatalii 0.9.0 JA	1/2	+0	2 Games
Stro4k 3.0 JA	1/2	+0	2 Games

Eubos 4.0 download