English Chess Forum

What is "neural net" and "github"

I often think this forum is a github.

Github is a subsidiary of Microsoft that my son works for.

JustinHorton wrote: ↑

Sat Feb 20, 2021 10:55 pm

What is "neural net" and "github"

Developers of "free" and open source software might use GitHub as the public storehouse of their shareable code.

Chess engines used to have humanly intuitive, fixed internal settings (like bonuses for the bishop pair or fewer pawn islands) whose values would have been optimised by endless testing and automated game-playing.

In neural networks the settings are no longer humanly intuitive or even describable. Computers are shown millions of positions and assessments, from which they derive their own settings and values, so to speak. These derived "networks" are opaque to us; we cannot fine tune them; we can only test one network against another. In neural networks the computer has learnt for itself what works in chess and represents this to itself quasi-mathematically, not conceptually.

Angus French wrote: ↑

Sat Feb 20, 2021 12:52 pm

The Lichess statement also states that FF2 was advertised as being the strongest engine available and in support of this presented results which were against not the version of Stockfish on which FF2 was based but an earlier version.

Chessbase now claim Fat Fritz scores 52.24% against Stockfish 13, giving it a 16 points rating advantage. CCRL agree it's better at blitz time controls, but only by 2 points. At a rapid time limit they have it 29 points worse.

Why anyone using one of these engines for home analysis, or to play against, would care about such tiny differences is beyond me.

....thanks to BREXIT I did get my copy tax free!!

I think the editorial style and bombast surrounding chessbase's claims isn't going to do them any favours in the long term.

From what I can gather Fat Fritz is based on an amalgamation of stockfish search + lc0 evaluation algorithms (as is the latest version of stockfish as per the lichess article) - the main difference being that Albert has trained these algorithms on a different net. Why this should make a tangible difference to the end user is unclear to me unless there is a clear difference in playing strength - but the evidence for this appears to be flimsy at best.

As a consumer I would have appreciated an approach that gave more credit to the original developers, explained the tangible differences caused by the different nets (for instance by discussing some differences in evaluation between different positions) and pointed out to customers that they can obtain stockfish + lco for free - whilst of course explaining the benefits of their approach and justifying the fee. For your 99 euro you are getting more than just the engine and it could well be that as a package this represents good value for money - but I'm disappointed by the lack of disclosure.

Alex Longson wrote: ↑

Mon Feb 22, 2021 4:24 pm

As a consumer I would have appreciated an approach that ... explained the tangible differences caused by the different nets (for instance by discussing some differences in evaluation between different positions)

My guess is that there aren't any, and, even if there were, who's to say which evaluation is better? I suspect that if you wanted to say one was better than the other you'd more likely be saying that they both agreed on the best move, but one found it a lot quicker than the other one did. ChessBase has previously used that claim in their marketing of other engines they sold when comparing a new engine to old engines.

Pete Doggers

The development teams behind the two most successful and influential open-source chess programs, Stockfish and Leela Chess Zero, have issued statements denouncing the commercial program Fat Fritz 2 and the company Chessbase that is selling the program for 99.90 euros

Here we go :


https://www.youtube.com/watch?v=9C-E1BqzLcs

Enjoy!

"Enjoy!"

Later perhaps - I find it interesting that comments are turned off. Maybe a wise precaution...

ChessBase has found someone who's willing to write an article claiming Fat Fritz 2 is better than Stockfish, primarily based on the author's assessment of one position.

He appears to be exclusively a correspondence chess player. No FIDE rating and no USCF rated games for 30 years.

One comment he made interested me -

fully neural net engines develop their nets by playing complete games. That experience and data is retained by their nets, enabling them to quickly and accurately play openings.

Is it the case that if I let Stockfish NNUE analyse one of my games it will record something in its neural net in case the same positions, or themes, arise again? If it does, it's not obvious. Last weekend I analysed my most recent game with Stockfish with a small hash table that was often filling completely. I then decided to do it again with a bigger hash table to see what difference that made. The analysis was slightly different in places, but there was no obvious speeding up where it was the same.

I think the issue here is why is so little or nothing said over how to use an engine and how they are being used? Reliance on technology isn't necessarily a sign of progress.

Haven't the times changed. With so many videos online it could be argued that chess literature is rapidly becoming defunkt and dinosaurish. And with playing platforms providing powerful engines, have the days where we need to buy such things has come and gone. What exactly are you profiting from by buying this stuff?

Ian Thompson wrote: ↑

Sat Feb 27, 2021 11:11 am

Is it the case that if I let Stockfish NNUE analyse one of my games it will record something in its neural net in case the same positions, or themes, arise again? If it does, it's not obvious. Last weekend I analysed my most recent game with Stockfish with a small hash table that was often filling completely. I then decided to do it again with a bigger hash table to see what difference that made. The analysis was slightly different in places, but there was no obvious speeding up where it was the same.

No. The training process is vastly more computationally expensive than actually running the neural nets produced. Basically done once on a big server or similar rather than continuous.

What gets shipped will be constant absent upgrades or very specific decisions.

The nice thing in this case is that it is a victimless crime, as all parties involved are responsible for causing irreparable damage to the game I am happy to see all of them lose out.

You do have to wonder what sort of people are playing actual money for this "fat Fritz" if it is based on stolen code from two FREE engines, especially given that programmers that operate in this way rarely improve on their source material.