Chess2u
Would you like to react to this message? Create an account in a few clicks or log in to continue.

Chess2uLog in

descriptionStosckfish LC0 hybrid EmptyStosckfish LC0 hybrid

more_horiz
Interesting combination of Stockfish and LC0:

https://github.com/benblack769/lc0-stockfish-hybrid

Does anyone know how to compile this?
Will there be updates?

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
This project is years old and never really got off the ground. It predates NNUE which allows most conventional engines to run neural networks.

Modern editions of Stockfish often use leela data for training nets, so they already play a bit like LC0.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
Stockfish 14 is a mix of Lc0 and SF Dev

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
Stockfish's Latest NNUE Architecture Is Mostly Made Of Leela Data
Stockfish Team Has Said That They Have A Collaborated With Leela Team To Make A New Network Architecture And Make The Engine Play With More Positional And In A Different Style
So nowadays They Train Nets Using Leela Data
Thank you mate

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
I know, LC0 data is used for training, but Stockfish does not use the GPU. If this engine runs on CPU and CPU at the same time, it is novel, even if out of date. Maybe updating it with the new Stockfish code would suffice?

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
@Jaglo wrote:
I know, LC0 data is used for training, but Stockfish does not use the GPU. If this engine runs on CPU and CPU at the same time, it is novel, even if out of date. Maybe updating it with the new Stockfish code would suffice?


We can try like that
But I dont know whether it would work or not
because the codes or very old and nearly 3years old
So the chances of making it is low
But trying is not wrong
We can try doing like that!

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
Could someone please prepare an exe file?

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
I am the author of this project. If people are interested in it, I would be willing to go in and try to redo the project with recent versions of Stockfish and Lc0 and run some basic benchmarks. I also have some ideas on how to improve it.

Unfortunately, it is not eligible for any of the major tournaments, so I lost interest in the project a few years ago.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
If I may ask, what is your interest in the project?

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
I think simultaneous computation on the processor and the graphics card is the future. This project seems to me to be the most advanced in this regard.

I am waiting for an engine of this type to appear. Whether it is qualified for tournaments or not is not important to me. His style of play will probably be superior to what we see so far. It's a milestone, even if it doesn't beat Stockfish in testing.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
Cool. At the time I made this project, there were a number of similar Lc0-stockfish style projects, but they did not make very good use of the unique qualities CPU resources I think, because they just didn't evaluate many nodes. After NNUE came out, I think everyone gave up on the idea of combining the projects, because they already seemed so similar.

I can link you to the current binary (over email) and see how much work it would be to update it to use more modern versions of the engine.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
So unfortunately, the answer appears to be: This is really hard. Massive merge conflicts with master branch of Lc0, and its hard to even track down the changes I made to the stockfish code.

This is several full days of work. So it might be awhile before I can get around to it. But I am interested in getting this project in a state which can be more easily benchmarked, updated, and iterated on. So I'll try to get around to it eventually.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
My first tests

1) LC0_unsolved test (10 AH_Tactics-250, 11 eigenmann_endgame_test)

LC0 0.21.2 (net 32988) - 0/21
Stockfish 10 64 POPCNT - 21/21
Hybrid - 16/21

2) Stockfish_unsolved test (6 openings, 6 gcp_positionaltest, 6 typp, 2 Karpov_games)

Stockfish 10 64 POPCNT - 0/20
LC0 0.20.1 (net 32988) - 20/20
Hybrid - 14/20

Very promising results

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
Wow, I didn't know there were known lists of problems that stockfish and lc0 could not solve. Where did you find those?

And I actually did find my old benchmark data which I said I deleted. Just lost on an obscure archive on a spare hard drive. So I'll put that up on my blog post and github readme soon.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
These are my private tests. I used publicly available tests and selected positions from them that one chess engine solved and the other did not. Stockfish turns out to be better than LC0 in tactical tasks and endgame. LC0 is good in positional play.

The above results show that the hybrid engine performs as one would expect. It can solve both tactical, endgame and positional problems. Obviously Stockfish 14 NNUE plays better than Stockfish 10. But my point was to see if the way the two chess engines are combined gives an advantage such that a more universal engine is created.

descriptionStosckfish LC0 hybrid EmptyRe: Stosckfish LC0 hybrid

more_horiz
Permissions in this forum:
You cannot reply to topics in this forum