StockFiNN Release and Scaling

[dkappe]

Is that actually the case, or are we simply getting away with it?

I’m just training my first NNUE net now, but generating useful data and training the comparatively small net seems to take at least one order of magnitude less time, maybe more.

[kranium]

It looks like testing and training are much shorter for NNUE, which may be to its advantage.

it's efficiently updateable

[Chessqueen]

What's the point of yet another fork of SF-NNUE? Any actual changes in NN or search to accommodate it, or TM?

The engine fork is not the interesting part. The interesting part is that I have trained a net that is as strong as Stockfish itself in a short TC 1 thread, and very rapidly without too many resources. And still gaining a lot

What is so special about StockFINN? As you can see here on TCEC LIVE https://tcec-chess.com/ It is no better than Stockfish, or LCZero, and it is slightly losing to Alliestien, the only engine that it can beat easy is Stoofvless, that is the only reason for a Plus Score so far, if you take Stoofvlees out, it would be losing

[MMarco]

What's the point of yet another fork of SF-NNUE? Any actual changes in NN or search to accommodate it, or TM?

The engine fork is not the interesting part. The interesting part is that I have trained a net that is as strong as Stockfish itself in a short TC 1 thread, and very rapidly without too many resources. And still gaining a lot

What is so special about StockFINN? As you can see here on TCEC LIVE https://tcec-chess.com/ It is no better than Stockfish, or LCZero, and it is slightly losing to Alliestien, the only engine that it can beat easy is Stoofvless, that is the only reason for a Plus Score so far, if you take Stoofvlees out, it would be losing

Lc0 and Stockfish are no better than StockFiNN, not to mention Stoovlees. Why bother to watch TCEC at all??

[FormazChar]

Stockfish dev ended up with -3 or-2 with same book. At these one ply openings and tcec hw&tc a extremly high draw is expected so all of the top tier engines will look close. The TCEC match shows that nnue fish is already close too normal sf. The hope is then that a lot of further gains will be made in the near future.

[Gejsi Marku]

It is no better than Stockfish

And SF doesn't seem superior to it either. That alone makes it speciall.

It is no better LCZero

Same reason as above.

and it is slightly losing to Alliestien

Wow.. such large sample size to agree with you. I guess Allie is better than Leela ans Sf based on what's going on on TCEC so far.

the only engine that it can beat easy is Stoofvless

Mind you two wins for SFNN came as black. Even top engines would have to really try hard against Stoof to do that.

[jjoshua2]

The main point is that it produces evals and moves more unique as a 2nd opinion to Stockfish than any stockfish clone or even Komodo, and yet manages to be stronger be much stronger than Komodo. Whether it is 30 elo above SF or below is largely irrelevant here. But there are some circumstances in which it has greater than 99% chance of being better. Even if it is equal to lc0 and produces similar evals and moves that is amazing because now you don't need an expensive GPU to play competitive ICCF for example.

[kranium]

What is so special about StockFINN? As you can see here on TCEC LIVE https://tcec-chess.com/ It is no better than Stockfish, or LCZero, and it is slightly losing to Alliestien, the only engine that it can beat easy is Stoofvless, that is the only reason for a Plus Score so far, if you take Stoofvlees out, it would be losing

The software's 1st release was on 2020-05-28...only about 6 weeks ago!

The 1st time I downloaded it (2020-6-09), I tried creating a net following the very basic steps outlined in the readme.txt.
It took all/most of 1-2 days to produce on my PC.

The result of initial testing...126 games, was (+23,=97,-6) for nnue.
http://talkchess.com/forum3/viewtopic.p ... ty#p850536

The point is: you can fairly easily, and very quickly train and produce an original evaluation for SF that equals or exceeds that newest SF dev version.
And this is just the beginning...imagine using more sophisticated training methods/techniques with a powerful server.
The potential is enormous.

Within it's 2nd month of existence, it has currently scored 3 wins and one loss out of 30 games...
vs the strongest group of chess engines SF, LCO, etc. that ever existed
https://tcec-chess.com/

This is pretty impressive stuff...

[jjoshua2]

People are understandably jaded from many bad SF clones that pretended to be stronger and very different from SF even with alphazero technology, but were in fact almost identical to SF... Makes it hard to convince people when in fact the entire evaluation function has no shared essence at all.

[carldaman]

If you try to select PayPal as a payment method on patreon it says it can't connect to it.