This version includes:
- a neural network trained "from zero", using only game results (1-0, 0-1, 1/2-1/2) and material as labels and targets ;
- quantization of weights and biases, resulting in a 40% increase in nps (nodes per second) !
- a completely rewritten Cerebrum library, allowing anyone to reproduce my work and obtain exactly the same neural network from the same data ;
- a minor change, with a default Transposition Table size of 256 Mb (previously 128 Mb).
While the new version may be weaker than the previous one (approximately 100 elo), what mattered to me was no longer relying on evaluations from other engines for training data. I'm really happy to get such decent performance in a original and reproductible way !
I also put a lot of efforts on the Cerebrum library:
- you now only need to provide PGN file(s) - or a handcrafted set of positions - and launch 2 scripts : everything is automated !
- you can customize the network’s name and the author’s name embedded in the network
- additionally, a basic UCI chess engine called “Cerebrum” is provided to demonstrate how to load and use trained networks in a real engine (Python)
Please note that Orion's URL has changed a few months ago and is now https://www.orionchess.com/
Feel free to share your feedback !
Kind regards from France,
David
