https://github.com/tomitank/tomitankChess/releases/tag/v5.0


This is quite different from the SF solution.

network is only 768x16x1. Smaller than I've seen before.

it does not replace evaluation, it only compensates that. (*)

trained only with 2.7M example.

The network is not compatible with other engines (*) and it's integrated into the code. (hardcoded)

I wrote the network from scratch. I don't use machine learning platform.

I have around 10 elo better net (768x32x1) in fix depth test, but JavaScript is too slow, and don’t have AVX or similar methode in JavaScript (so i use vanilla approach). Essentially the smaller net has nearly the same strength in both short and long TC tests.

The training method is similar to that described by Andrew, thanks to him for the explanation.
But the implementation is unique and I have not received any help.
http://talkchess.com/forum3/viewtopic.p ... 72#p874960

This net is not perfect, and HCE slows that down further, but i win around 60-70 or even more. (i don't tested deeply.) In C it would be around 140-150.

I can try to use more examples later, but it’s pretty boring for me.

I also plan to turn off HCE and use only NN, but (probably) the authors of Halogen do this better.

I tried to go my own way, so feel free to send this version to the CCRL list.

Estimated elo: near to 2900.

Since JavaScript is about 4-5x slower than C, it would be around 3100 or above in C. (With AVX).

JavaScript uses only one core. With 4 cores, in C it would be near to 3300.

I did not tune the search. It could probably be even stronger.

Have fun with tomitankChess!