The Next very strong Engine will be rofChade NNUE

[Chessqueen]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

[AndrewGrant]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

[MikeGL]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

I agree. But my observation on some test sets (see my marshall thread on Nf4) shows that Komodo Dragon net is so strong and well rounded compared to any other nets.

[Chessqueen]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

You stated here that Ethereal SF NNUE gained +172 "Ethereal 12.75 SF-NNUE +172" , but How much do you believe rofChade can gain without using SF NNUE, but by using its own NET, Probably 75 Elo

[Chessqueen]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

I guess it is possible to get +150 without using SF NNUE Net

Re: New release Pedone 3.0
Unread post by pohl4711 » Wed Jan 06, 2021 6:05 am

Fabio Gobbato wrote: ↑Fri Jan 01, 2021 5:23 am
I've just released a new version of my engine.

These are the changes:
- Pure neural network evaluation
- Tuning search parameters
- Fix a bug in chess960 mode
- Improved futility pruning formula

Lets say that from the changes listed above Pedone gained 75 Elo, and from the Pedone NET another 75 Elo, therefore, rofChade could gain at least 75 from a rofChade NET

For this release I would like to thank Dale Weiler for the header incbin used to embed the neural network in the executable.
Compared to version 2.1 is about 150 elo stronger.
I have removed from the package the 32 bit executables.

The download is from Pedone site: https://sites.google.com/site/pedonechess
7000 games testrun of Pedone 3 avx2 finished:

https://www.sp-cc.de/

[AndrewGrant]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

I guess it is possible to get +150 without using SF NNUE Net

I'm referring to top level engines. Pedone is no where near a top level engine.

[Graham Banks]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

I guess it is possible to get +150 without using SF NNUE Net

I'm referring to top level engines. Pedone is no where near a top level engine.

Possibly only 100 Elo behind Ethereal 12.75.

[AndrewGrant]

Possibly only 100 Elo behind Ethereal 12.75.

Well so more generally this opens the question about Network training, and what amount of difficulty is required.

Right now, +70 from 12.75 is the best I can do, after playing around ~40,000,000 games, training on a few billion positions, using a machine with 104 threads. Noob basically gave me short term access to more computational power than most people can even begin to muster for such a trial and error task like this.

Now you can do better than my +70 (2x128 nets) by using a 2x256 network if you copy paste SF's trainer (and implementation ?) because the speed is similar -- the crucial difference is weight quality. Compressing stuff down to 8 bits seems very hard. I use 16 bits for two layers, and floats for the rest. Using 16 bits for final layers kills elo. Its simply very hard to compress to perform that level of quantization.

I'm not completely retarded, but I did do this all myself. So its possible I have issues in the training code. Less so that there are issues in the implementation code. So its not like my results are a highbar in terms of what is possible. But it is a bar in terms of what is doable by one man working on his own.

[Chessqueen]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

I guess it is possible to get +150 without using SF NNUE Net

Re: New release Pedone 3.0
Unread post by pohl4711 » Wed Jan 06, 2021 6:05 am

Fabio Gobbato wrote: ↑Fri Jan 01, 2021 5:23 am
I've just released a new version of my engine.

These are the changes:
- Pure neural network evaluation
- Tuning search parameters
- Fix a bug in chess960 mode
- Improved futility pruning formula

Lets say that from the changes listed above Pedone gained 80 Elo, and from the Pedone NET another 70 Elo, therefore, rofChade could gain at least 70 from a rofChade NET, plus I believe that Komodo also gained at least 80 using Komodo NET

For this release I would like to thank Dale Weiler for the header incbin used to embed the neural network in the executable.
Compared to version 2.1 is about 150 elo stronger.
I have removed from the package the 32 bit executables.

The download is from Pedone site: https://sites.google.com/site/pedonechess
7000 games testrun of Pedone 3 avx2 finished:

https://www.sp-cc.de/

[Chessqueen]

At the moment rofChade is NOT too far away from Ethereal in TCEC, but if NNUE give rofChade an extra 150 Elo it will be on the Top 4
https://tcec-chess.com/live.html

Only way you get +150 is by taking an SF network

I guess it is possible to get +150 without using SF NNUE Net

Re: New release Pedone 3.0
Unread post by pohl4711 » Wed Jan 06, 2021 6:05 am

Fabio Gobbato wrote: ↑Fri Jan 01, 2021 5:23 am
I've just released a new version of my engine.

These are the changes:
- Pure neural network evaluation
- Tuning search parameters
- Fix a bug in chess960 mode
- Improved futility pruning formula

Lets say that from the changes listed above Pedone gained 80 Elo, and from the Pedone NET another 70 Elo, therefore, rofChade could gain at least 70 from a rofChade NET, plus I believe that Komodo also gained at least 73 using Komodo NET. At the moment Ethereal is only about 62+ Elo stronger than rofChade, so roChade with a NET could easily reach the same rating of Ethereal

Ethereal
TCEC S20 DivP
34 12.5 36.76% 3 [3/0] 12 [0/12] 19 [15/4] 12.629 [0.129] 207.75 3482 -8 [3474]
8
rofChade
2.309
34 9.5 27.94% 2 [2/0] 17 [1/16] 15 [13/2] 9.625 [0.125] 168 3423 -13 [3410]