SF-NNUE going forward...

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, Dann Corbit, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
peter
Posts: 2171
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: SF-NNUE going forward...

Post by peter » Wed Jul 29, 2020 6:12 am

Ovyron wrote:
Wed Jul 29, 2020 2:43 am
[Black "Ulycrom, Chi-chi+NNUE"]
That's not fair.
:)
You won't make it public, will you?
:twisted:
Peter.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Wed Jul 29, 2020 6:18 am

peter wrote:
Wed Jul 29, 2020 6:12 am
That's not fair.
:)
You won't make it public, will you?
:twisted:
Huh, I guess I'll make it public to people that promise to make personalities with it. It's just this one with tweakable settings.

peter
Posts: 2171
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: SF-NNUE going forward...

Post by peter » Wed Jul 29, 2020 6:21 am

Ovyron wrote:
Wed Jul 29, 2020 6:18 am
peter wrote:
Wed Jul 29, 2020 6:12 am
That's not fair.
:)
You won't make it public, will you?
:twisted:
Huh, I guess I'll make it public to people that promise to make personalities with it. It's just this one with tweakable settings.
Of course I know ChiChi already, and of course I did fiddle around with it, and of course I promise to make personalities with it even more with NNUE!
:)
Seriously: wouldn't it be a fine idea too, to train some nets of their own with some personality- evals for start of learning instead of normal SF- evals?

Hope dkappe reads this.
:)
Peter.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Wed Jul 29, 2020 7:50 am

peter wrote:
Wed Jul 29, 2020 6:21 am
Of course I know ChiChi already, and of course I did fiddle around with it, and of course I promise to make personalities with it even more with NNUE!
:)
Seriously: wouldn't it be a fine idea too, to train some nets of their own with some personality- evals for start of learning instead of normal SF- evals?

Hope dkappe reads this.
:)
LOL! As I went to release Chi-chi NNUE I tested it to see if it worked... and it didn't! Changing the piece values and such doesn't do anything! They're just decorative...

Now I wonder if it could be done at all, or if Stockfish-NNUE is just bypassing all eval terms and just evaluating directly from the network, which would make Chi-chi obsolete before release :shock:

User avatar
towforce
Posts: 10694
Joined: Wed Mar 08, 2006 11:57 pm
Location: Birmingham UK

Re: SF-NNUE going forward...

Post by towforce » Wed Jul 29, 2020 7:52 am

Dann Corbit wrote:
Wed Jul 29, 2020 5:54 am
Ovyron wrote:
Tue Jul 28, 2020 8:22 pm
marsell wrote:
Tue Jul 28, 2020 9:01 am
People need several expensive and power-hungry GPUs to beat Stockfish
What is interesting is that those people got the GPUs for Leela. They still have the GPUs. The GPUs didn't just disappear.
GPUs are marvelous, incredible things. I bought GPUs long before I found out they can be useful for chess. I have GPUs in all of my machines and boxes of GPUs besides.

You can test Mersenne Prime candidates.
You can fold proteins and cure diseases.
You can look for gravitational waves
It's not a case of what can you do with a GPU, it's what can't you do. I guess you can even stop a herd of stampeding Yak with GPUs.

The uses are endless. I think GPUs are the most fabulous thing (hardware-wise) on the surface of the earth. I can't wait to buy some more and some really smoking hot ones are just around the corner

19 years ago, the world's fastest computer (the computer that topped the Top500 chart) was ASCI White. It weighed 106 tons and needed 3 MW of power. Now you can get the same processing power from the Nvidia RTX 2080 Ti graphics card.

In terms of FLOPS, you can probably get even more for cost/power usage from TPU chips (tensor processing unit, designed for machine learning (ML)), but these devices mostly use half-precision arithmetic, which limits their usage outside of the ML domain. The fastest computer I'm aware of that you can order today and have delivered to your home in couple of days is the Nvidia DGX A100, which has a 5 petaflop performance (Kurzweil's estimate of the human brain is 20 petaflops), but I think that most of this is half precision, so again limited usage outside of ML.

Because you can rent time on Google TPUs on Google Compute Engine (GCE), GCE has a higher top speed than any of its cloud rivals.
Writing is the antidote to confusion

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Wed Jul 29, 2020 8:01 am

Zenmastur wrote:
Mon Jul 27, 2020 6:59 pm
Ovyron wrote:
Mon Jul 27, 2020 5:35 am
We're currently in a crisis at which 60% performance against Stockfish 11 hasn't been achieved by any net, if more positions don't help and more depth doesn't help, and bigger net doesn't help, then, yeah, let's start trying other things like multiple nets for game stage.
A crisis???

A little melodramatic don't you think?
Update: already at 64% and still going up.

So, huh, well, Sergio's approach keeps on working, at least for now the recipe for success is: train, train, train. And then train some more. No saturation in the horizon.

Things are moving so fast that whatever I say may be irrelevant 2 days from now, if when it was released someone had told me Stockfish 11 was going to be destroyed like this in July I'd have called a mental institution.

peter
Posts: 2171
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: SF-NNUE going forward...

Post by peter » Wed Jul 29, 2020 8:07 am

Ovyron wrote:
Wed Jul 29, 2020 7:50 am
LOL! As I went to release Chi-chi NNUE I tested it to see if it worked... and it didn't! Changing the piece values and such doesn't do anything! They're just decorative...

Now I wonder if it could be done at all, or if Stockfish-NNUE is just bypassing all eval terms and just evaluating directly from the network, which would make Chi-chi obsolete before release :shock:
It's doeing something for sure to change material- and or positional evals as fo ChiChi itself (without NNUE).
Probably it doesn't togehter with NNUE, because hard coded evals are replaced by NNUE- ones with net loaded, Contempt doesn't do a thing as for SF NNUE neither.

Maybe as for ChiChi "pure", if you don't change all of the material ones together, it give's even more effect of will to sacrifice figures. My best personality so far is one with doubled pawn- values and halved the rest of the material ones. That togehter with raised King Safety get's a touch of legendary Aggressor Dissident made out of Zappa by Robert Flesher, you remember this setting?

Give it a try, if you like to.
:twisted:

And as for special nets, dkappe did a training with evals of komodo e.g. for his lizzard- fish and with Toga - evals. Giving a ChiChi personality a try with changed evals of SF as starting point would be fun too indeed maybe. Pity I'm not skilled enough in creating NNUE nets so far to do it on my own.
But I'l tell Dietrich about it at CSS maybe, if it's ok with you.
:?:
Peter.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Wed Jul 29, 2020 8:29 pm

peter wrote:
Wed Jul 29, 2020 8:07 am
Contempt doesn't do a thing as for SF NNUE neither.
Oh, so NNUEs are an all or nothing deal.
peter wrote:
Wed Jul 29, 2020 8:07 am
My best personality so far is one with doubled pawn- values and halved the rest of the material ones. That togehter with raised King Safety get's a touch of legendary Aggressor Dissident made out of Zappa by Robert Flesher, you remember this setting?

Give it a try, if you like to.
:twisted:
Yes! I remember. I'll do it, it'll be interesting to see how low Chi-chi's depth needs to be to allow Aggressor Dissident to catch up 8-)
peter wrote:
Wed Jul 29, 2020 8:07 am
But I'l tell Dietrich about it at CSS maybe, if it's ok with you.
:?:
Sure, but maybe wait till my next Chi-chi release which will include material asymmetry tweaking (to allow the engine to think that some opponent's pieces are more or less valuable that your own, or vice versa, ala Rybka tweaking.)

peter
Posts: 2171
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: SF-NNUE going forward...

Post by peter » Wed Jul 29, 2020 10:29 pm

Ovyron wrote:
Wed Jul 29, 2020 8:29 pm
Sure, but maybe wait till my next Chi-chi release which will include material asymmetry tweaking (to allow the engine to think that some opponent's pieces are more or less valuable that your own, or vice versa, ala Rybka tweaking.)
No hurry, guess dkappe will have read about ChiChi already anyhow.
Last edited by peter on Wed Jul 29, 2020 10:31 pm, edited 2 times in total.
Peter.

ChickenLogic
Posts: 69
Joined: Sun Jan 20, 2019 10:23 am
Full name: Julian Willemer

Re: SF-NNUE going forward...

Post by ChickenLogic » Wed Jul 29, 2020 10:30 pm

Haven't read all of this thread so excuse me if it was already mentioned: tttak's fork has a lot of different architectures to try out. I'm interested in the "Gameply40x4" one which effectively holds 4 NNs in one file which get used sequentially switching every 40 plies until the 4th is loaded. This sacrifices no speed over the regular halfkp-nets with the same input layer size. It'll require at least 4x more training data though. With weaker nets this will introduce more eval instability of course but I think a "more correct" eval will benefit search more than eval instability will hurt it. Maybe it is also possible to make NNUE viable on really old hw with "Gameply40x4" nets with smaller input layers.
I think this concept of multiple nets will work better for SF than for Leela. The most obvious reason is that it takes less positions to saturate NNUE nets and thus we can experiment much more.
Currently working on SFNN.

Post Reply