NNUE
-
- Posts: 20
- Joined: Mon Oct 17, 2016 09:05
- Real name: Robin Messemer
Re: NNUE
How reasonable would it be to train a nnue-policy-network to use for move-ordering ? Has anyone already tried that ?
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
Hi Joost.Joost Buijs wrote: ↑Sat Apr 10, 2021 18:34It looks as if the network file is not loaded. Scan-NNUE doesn't seem to check if the file is present or not.
The file nn_20210328.gnn should be in the data folder. I had the same issue here too.
Well I guess you are right. Anyway, I don't know what to think about it anymore.
-
- Posts: 471
- Joined: Wed May 04, 2016 11:45
- Real name: Joost Buijs
Re: NNUE
Well, it just works. Two days ago I tried a match with it against KR, single core, 90 moves per minute, it came out at -13 Elo. With very fast games it is still at a disadvantage, mainly because of the slower speed. With games of 75 moves in 3 minutes it is about equal. You need a computer with fast AVX2 to get the most of it.Krzysztof Grzelak wrote: ↑Sat Apr 10, 2021 18:59Hi Joost.Joost Buijs wrote: ↑Sat Apr 10, 2021 18:34It looks as if the network file is not loaded. Scan-NNUE doesn't seem to check if the file is present or not.
The file nn_20210328.gnn should be in the data folder. I had the same issue here too.
Well I guess you are right. Anyway, I don't know what to think about it anymore.
In the future it will only get better, better hardware for inference, better networks. It took several years before pattern evaluation was on the level it is now, for these NNUE style nets it won't be any different.
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
I think you don't need better processors these days. It is better to make a program, even if such a Stockfish.Joost Buijs wrote: ↑Sat Apr 10, 2021 19:19Hi Joost.
Well, it just works. Two days ago I tried a match with it against KR, single core, 90 moves per minute, it came out at -13 Elo. With very fast games it is still at a disadvantage, mainly because of the slower speed. With games of 75 moves in 3 minutes it is about equal. You need a computer with fast AVX2 to get the most of it.
In the future it will only get better, better hardware for inference, better networks. It took several years before pattern evaluation was on the level it is now, for these NNUE style nets it won't be any different.
-
- Posts: 471
- Joined: Wed May 04, 2016 11:45
- Real name: Joost Buijs
Re: NNUE
I dont like Stockfish, I'd rather have a nice Dutch salted herring!Krzysztof Grzelak wrote: ↑Sat Apr 10, 2021 20:10I think you don't need better processors these days. It is better to make a program, even if such a Stockfish.
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
Kingsrow - Scan 3.1 NNUE - TC - 1 minute 75 moves, 6P EGDB, book - off, pondering - off, threads -1. Result of the match - 9 wins, 0 losses, 147 draws, 2 unknowns
- Attachments
-
- dxpgames.pdn
- (151.41 KiB) Downloaded 330 times
-
- Posts: 471
- Joined: Wed May 04, 2016 11:45
- Real name: Joost Buijs
Re: NNUE
This is almost the same as what I got over here:
Kingsrow 1.62 vs Scan_nnue 3.1: 7 wins, 1 losses, 150 draws, 0 unknowns
Conditions: 1 minute 90 moves, 1 thread, no pondering, book off, 6p EGDB.
With 1 thread Scan-NNUE does 4.5 mnps on my (not overclocked) 3800 MHz. i9-10980XE. I haven't tried on the AMD thread-ripper yet.
Kingsrow 1.62 vs Scan_nnue 3.1: 7 wins, 1 losses, 150 draws, 0 unknowns
Conditions: 1 minute 90 moves, 1 thread, no pondering, book off, 6p EGDB.
With 1 thread Scan-NNUE does 4.5 mnps on my (not overclocked) 3800 MHz. i9-10980XE. I haven't tried on the AMD thread-ripper yet.
- Attachments
-
- dxpgames-131.pdn
- Match result
- (152.23 KiB) Downloaded 339 times
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
I have such a question for Bert. And you can do thread=8
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
Thanks for the answer Bert.
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
Sorry to ask Bert - if the book is working properly.
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
Kingsrow 1.62 vs Scan_nnue 3.1: 4 wins, 1 losses, 153 draws, 0 unknowns
Conditions: 1 minute 90 moves, 1 thread, no pondering, book off, 6p EGDB.
Match played on a computer with the equipment.
Processor - AMD Ryzen Threadripper 1950X
Hard disc - Samsung 970 EVO Plus NVMe M.2 SSD 2 TB
Memory of frames - 128 GB DDR4 2400
System - Windows 10 64 bit Pro
Conditions: 1 minute 90 moves, 1 thread, no pondering, book off, 6p EGDB.
Match played on a computer with the equipment.
Processor - AMD Ryzen Threadripper 1950X
Hard disc - Samsung 970 EVO Plus NVMe M.2 SSD 2 TB
Memory of frames - 128 GB DDR4 2400
System - Windows 10 64 bit Pro
- Attachments
-
- dxpgames.pdn
- (151 KiB) Downloaded 314 times
-
- Posts: 1368
- Joined: Thu Jun 20, 2013 17:16
- Real name: Krzysztof Grzelak
Re: NNUE
How reasonable would it be to train a nnue-policy-network to use for move-ordering ? Has anyone already tried that ?
So far as I know no-one has tried this for draughts (or checkers).
But I'm convinced it can be done.
As the capture (and kings) add so many moves to the moves-set, maybe on need to focus (only) on static man-moves.
Whether the improved moved ordering offset the time-penalty for the nnue calculation and the reduced search depth, and if this method yields better results then history (which I consider often as random generator ) remains to be seen.
But certainly something worth trying in the future.
And if the network is perfect, then you at least don't need search anymore.
For the short term, my focus at least, is improving nnue, and try to close the gap with Scan and Kingsrow.
Bert
So far as I know no-one has tried this for draughts (or checkers).
But I'm convinced it can be done.
As the capture (and kings) add so many moves to the moves-set, maybe on need to focus (only) on static man-moves.
Whether the improved moved ordering offset the time-penalty for the nnue calculation and the reduced search depth, and if this method yields better results then history (which I consider often as random generator ) remains to be seen.
But certainly something worth trying in the future.
And if the network is perfect, then you at least don't need search anymore.
For the short term, my focus at least, is improving nnue, and try to close the gap with Scan and Kingsrow.
Bert