Is Talkchess still the centre of computer chess programming?

[Madeleine Birchfield]

Frank, your discussion about chess GUIs and neural networks is off topic. I kindly ask you to start a new discussion about that.

[Frank Quisinsky]

Madeleine, you discussion about the same topic again and again in most of your TalkChess messages is very boring for myself!
I try to make it more interesting!

Your question:
Thoughts? Is Talkchess still the premier place to go to talk about computer chess programming, or are there alternatives to Talkchess that are better for computer chess programming discussions?

I gave an answer!

[hgm]

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.

You cannot expect there to be a discussion between engine authors if there is only one engine. And there only is Leela, and it is just an implementation of what is described in the AlphaZero paper. What library to use is not a development issue; it has as little to do with Chess as what compiler would give the fastest Stockfish executable. When a few users that know how to compile discuss such compilation matters, that does not count as a discussion between engine developers. No matter where you conduct it. There have been discussions here on what could be better NN topologies and better search algorithms. But the Leela programmers made it very clear they were not interested in that, because it was not what the AlphaZero paper said, and they wanted to make an exact replica of AlphaZero. So it doesn't seem there is anything to discuss...

[KLc]

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.

Hm, then why not either contribute a discussion or simply use the other sources? As mentioned above already, talkchess is an incredibly valuable source of information dating way back.

[JohnWoe]

For starters NN engines like lc0 you need powerful GPU cards. I'm not investing 1200€ for RTX 3080 (+ more for the other stuff) just run some chess engine. Maybe mining to get my investment back. There are probably interesting chess topics in GPU based NN engines. But I'm not interested in chess point of view. Alpha-beta is 70 years old and serial by nature. Aka hard to parallelize search. But there's nothing better. Some new simpler parallel search algorithm would be nice.

When it comes to programming. I'm not posting questions to anywhere and wait hours for somebody to answer.

[towforce]

Thoughts? Is Talkchess still the premier place to go to talk about computer chess programming, or are there alternatives to Talkchess that are better for computer chess programming discussions?

Seems a little like asking, "Which is the best radio station to listen to while driving?"

I'm going to go "big picture" and briefly discuss what tends to happen to markets:

* early movers tend to get the biggest share of the market. This advantage, becoming the brand name, often lasts a hundred years or more in some business sectors

* companies that enter the market later have a huge problem with not being the brand name. The best adaptation is to create a new sector within the market, and make yourself the brand name in that new sector. This is why, in the 20th century, "the century of the song", we ended up with such a huge number of different types of music. IMO, places-for-computer-chess-discussion are heading in this direction at the moment

* when a new market entrant does manage to displace the brand name, in most cases, they did it via the "good enough and cheaper" route

* eventually you generally end up with too many suppliers to a market, and you get a "shakeout". IMO, computer chess has been in this state for a long time, and actually still is: there are huge numbers of people who want to write chess programs. The price is already zero dollars (except for a small number of brand names) and cannot go any lower. Based on historical trend, I think that enthusiasm for writing chess programs will eventually abate. When it goes to nearly zero (I don't think it will die completely - many formerly popular hobbies still have a few die-hard people still doing them), then I suspect that the number of places-for-computer-chess-discussion will also decline. I suspect that Talkchess won't be the last man standing because forum software like this is a bit outdated to be honest. Of course, it might adapt by becoming something like a Facebook group or whatever will be current in the future.

[Tony P.]

There have been discussions here on what could be better NN topologies and better search algorithms. But the Leela programmers made it very clear they were not interested in that, because it was not what the AlphaZero paper said, and they wanted to make an exact replica of AlphaZero.

This is no longer the case. There have been enhancements like squeeze-excitation and the moves-left head that A0 lacks. There's been a lot of experimentation with hyperparameters amid the training runs. Various net sizes have been tried. There's even more variety in training methods, giving rise to competing series of nets (T60 vs JH vs SV, etc.), and conversely, there are alternative engines using Leela nets. The exploration is limited by the huge length of the training cycle, which makes the project less resilient than the NNUE family, but it's plentiful anyway. What may ultimately bury the Leela project is the pricing of GPUs driven by the cryptocurrency market.

As for the original question: TalkChess is no longer the one and only source for CC knowledge, but it does have useful discussions of low-level routines like board representation and move generation. It's a complement to CPW and the SF and Leela Discord servers. It makes sense and is quite easy to read them all for matters specific to CC.

That said, CC has entered the era of machine learning, and to keep up to date with the latter, general scientific sources like Semantic Scholar are needed and may already require more time than CC sources in an ambitious engine dev workflow.

[smatovic]

>> Is Talkchess still the centre of computer chess programming?

I guess it never was, it was always a kind of meta-place, different kind of
breeds coming together...I value it.

--
Srdja

[Gabor Szots]

Rolf reincarnated.

[Ferdy]

Thoughts? Is Talkchess still the premier place to go to talk about computer chess programming, or are there alternatives to Talkchess that are better for computer chess programming discussions?

I believe it still is. I have not known any site that is better than talkchess.

Lc0 discord is of course better about Lc0 and mcts. Stockfish discord is good. Open chess is fine.