TalkChess.com

Now in the download site there is a graphic of rating (fixed d12 games).

https://www.comp.nus.edu.sg/~sergio-v/nnue/

Now the latest one 20200728-0633.bin seems the strongest, almost outside error margins again in 1000 games. Now there are substantial Elo gains versus old 2141, 10-12 Elo points or so. Fist I tested a bit the two new nets: 20+0.1 # PLAYER : RATING ERROR POINTS PLAYED (%) 1 stnnue20200728-1442 : 285...

The new one 20200727-0928.bin seems on a par with the best in 1000 games. Yes. It seems equal to 2051. 20+0.1: # PLAYER : RATING ERROR POINTS PLAYED (%) 1 stnnuesergio20200727-0928 : 2856.4 1.8 4629.5 9239 50.1% 2 stnnuesergio20200725-2051 : 2855.6 1.8 4609.5 9239 49.9% In 1000 games matches, today...

The new one 20200727-0928.bin seems on a par with the best in 1000 games. Yes. It seems equal to 2051. 20+0.1: # PLAYER : RATING ERROR POINTS PLAYED (%) 1 stnnuesergio20200727-0928 : 2856.4 1.8 4629.5 9239 50.1% 2 stnnuesergio20200725-2051 : 2855.6 1.8 4609.5 9239 49.9%

Tuning search is always fine, tuning position- learning code based on hash- learning wouldn't make code much more complicated, I think, even if I'm not good enough in proramming to prove so. Andscacs had a nice hash- storage before SugaR had so, no interest in position- learning based on selected h...

Tuning search is always fine, tuning position- learning code based on hash- learning wouldn't make code much more complicated, I think, even if I'm not good enough in proramming to prove so. Andscacs had a nice hash- storage before SugaR had so, no interest in position- learning based on selected h...

I also think that tuning search for the new NNUE eval will net nice redits. This will complicate the code of Stockfish, though.

2051 seems the best: Someone found at Discord that experimental-20200722-1621.bin is even better. Testing... 2051 prevails. 20+0.1 # PLAYER : RATING ERROR POINTS PLAYED (%) 1 stnnuesergio20200725-2051 : 2857.3 1.8 4659.5 9249 50.4% 2 stnnuesx20200722-1621 : 2854.7 1.8 4589.5 9249 49.6%

My first idea was specialized nets. For example a net for rook endgames. But the fact that the previous net that arrives to a rook endgame does not know about those endgames, will weaken it. So once you start using a net, I think the best is use it till the end. So maybe various nets specialized in ...

cdani wrote: ↑

Sun Jul 26, 2020 6:32 pm

2051 seems the best:

Someone found at Discord that experimental-20200722-1621.bin is even better. Testing...