Is Talkchess still the centre of computer chess programming?

[Madeleine Birchfield]

Are these programmers on TalkChess? I must have missed their postings, then...

BBC author: Maksim Korsh. Igel author: voffka. Rubichess author: RubiChess. Nemorino author: Florentino. Marvin author: Martin. All of them have been active members of Talkchess for the past few years and have posted and participated in the discussions about NNUE and various other computer chess programming discussions on Talkchess in the past six to nine months, both on the general subforum and in the programming subforum. A few others I should add are the likes of John Woe (Mayham/Sapeli), Daniel Shawul (Scorpio), dkappe (Toga III), and Engin (Tornado) to the list. I don't know where you were in the second half of 2020 to have missed all of this, maybe too busy trying to shut down the CTF.

[Madeleine Birchfield]

And what is bad if programmers "copy the nodchip NNUE architecture into their engine" ??

It is only bad if you are like H.G. Muller and want to see engines use a neural network architecture that is designed specifically for chess and not for shogi. I personally no longer care either way.

[hgm]

idea comes not from myself.
A good known programmer are thinking about it and sent me for three months a mail about it.
He like to contact Stefan Meyer-Kahlen or Martin Blume.

This is a totally non-sensical idea. An engine needs to consult its evaluation many millions of times per second. Doing that via the GUI woul slow it down by a factor 1000 or so.

[hgm]

It is only bad if you are like H.G. Muller and want to see engines use a neural network architecture that is designed specifically for chess and not for shogi. I personally no longer care either way.

Oh? Did you care before, and what made you change your mind?

It seems that what you care about isn't really that important, on the scale of things. I actually would think it is of exactly zero importance. What matters is if engine would gain a couple of hundred Elo by using a net architecture adapted to Chess.

[Ras]

Made sense what you wrote but the idea is to create NN files directly with the GUI

That's what my suggestion would allow - generating a generic NNUE file, and it would make NNUE generation independent from specific engines so that GUIs could be developed against that standard.

and engines can used the file via a special protocol.

That doesn't make sense. You need a protocol only if you have different programs communicating with each other, and you don't want the NNUE to be in a program other than the engine. The performance would be horrible. A protocol is not the right approach here - a format standard for NNUE data files is.

[Frank Quisinsky]

Hello HGM,

you are the programmer, you can evaluate the idea!

Why programmers should develop NN files themselfes?
Fully OK if programmers do that for more originialty but all the work around it ...

Often I am thinking, time can be better used!
So I like the idea but I am not a programmer and can evaluate the idea!

Best
Frank


Hello Ras,

thanks for your explanation.
I can understand it.

Best
Frank

[hgm]

I think you confuse training a net with developing one.

[Madeleine Birchfield]

Oh? Did you care before, and what made you change your mind?

It seems that what you care about isn't really that important, on the scale of things. I actually would think it is of exactly zero importance. What matters is if engine would gain a couple of hundred Elo by using a net architecture adapted to Chess.

It is precisely this that I no longer care about anymore. A few months ago I did, along with the likes of Andrew Grant, noobpwnftw, Connor McMonigle, Florentino, et cetera, and so it seems, you. In fact, I clearly remember a discussion in the programming subforum about the BONUS_PIECE_ZERO field in the nodchip NNUE architecture and how it was useful in shogi and crazyhouse but was useless in regular chess, which is why Florentino removed it from Nemorino.

The reason why I no longer care about it as much anymore is because I no longer care as much about CPU-only alpha-beta chess engines, especially after the scandals in the community regarding Houdini, Fire, Fat Fritz 2, Chessbase, and the rating lists.

[Madeleine Birchfield]

More comments from the other thread from somebody who missed my comment:

What exactly is the incentive to help people who come into the forum working on their first ever move generator? Obviously its a kind thing to do.

If someone who knows all the tricks of the trade thinks himself to be too good to help people just starting out, he's forgotten the fact that some time ago, he himself had to start somewhere.

To be honest, I could have had my engine in the 3000+ range already. I could have just written the move generator and alpha-beta functions, and then go around the stronger engines, taking the tapered evaluation from there, evaluation terms from here, and some prunings from one or the other search function. Easy enough to build a strong engine that way. Especially if I had chosen C or C++.

I didn't want to. I chose Rust, a programming language I didn't know a lot about when I started using it, and I chose to research the topics in chess programming from the very basics and write everything from scratch myself. Thus, at some points, I needed help.

A high-up-there discord server where people discuss the latest in neural networks is not the place to find information. A forum like this is; the 20-30 years of archived posts in forums like this, and newsgroups, is the place to find that information. Maybe, at some point in time, I'll end up on a discord server... but I doubt it. I do not have the time nor the motivation to start researching ways to gain 10 Elo and then play 50.000 games to see if it's actually a gain.

As soon as I feel my engine is going to take too much work and too much computing power to improve, I'd rather help other people getting started with THEIR first engine.

The old techniques won't go away: knowing how to write algorithms such as tree search are useful for many other games and applications. Writing this chess engine has taught mu much more than only chess programming, and gave me the chance to use and write things I'd not write during a normal working day.

Maybe you can argue it helps bring more people into the community, which is a net gain in the long run. But on an individual level, why should I, or a Stockfish dev, or someone else spend the time weeding through someone's first ever chess code with hundreds of lines to hunt down their bugs for them?

You don't, but you can at least point out the pitfalls which people need to take into account.

I think you missed his point. Top engine devs are not lurking around in the talkchess subforums anymore. They are in discord servers. Whether that be Stockfish, Leela, or smaller ones like their own Koivisto one or OpenBench's discord.

I think that's a pity. Discord is a chat program and the signal to noise ratio is usually very low. It is not a good place to preserve information (there are no threads for example). Sadly many people seem addicted to it.

This too. Where can you find 20-30 YEARS worth of information in Discord?

When googling for something, I invariably ended up in the CCC newsgroup, or on Talkchess, and except for a very minor instances, I found all the information I needed in sometimes VERY old posts.

All of this is only applicable for CPU-only alpha-beta engines. Talkchess is near useless for writing GPU-based MCTS-based engines with deep neural networks with 30-40 layers; there is very little discussion on Talkchess about CUDA vs Blas vs other backends, different linear or tensor algebra libraries to use, the different MCTS or UCT algorithms to use, et cetera, core foundational features of a Leela/AlphaZero style engine. For that, the resources are in the AlphaZero paper, the Leela discord, and various pages on github, medium, and so forth.

[Frank Quisinsky]

Hi HGM,

that's the point!
Training a net can be a great feature for a chess GUI with some nice possibilites for users of engines.

In my opinion programmers can use the time better as for "training a net".

What's your opinion about it?
You are working on the nice Winboard GUI and you made so many good things around Winboard.

Do you think that programmers should work on here engines and in additional on "training a net".
Two different things for me ... maybe better to isolate it?

I am not alone in thinking "confusing around nets".

Best
Frank