Book chapter on computer chess

[Kappatoo]

Hi all, I have a draft of a book chapter on game-playing AI, with a focus on computer chess.
The chapter goes through some history of computer chess, explains some basics about how chess engines work, and then describes what we have learned from computer chess.
If you are interested in reading the chapter and potentially giving me some feedback, and you also think you have relevant expertise in at least one of these topics, send me a message.
I would also be happy to discuss some aspects of the chapter here, if there is interest.

[jefk]

can't find you're real name, but give me a pm
and i see what i can do (if i can read the chapter)

[smatovic]

Maybe two related threads in context "what we have learned":

"The Bitter Lesson" by Richard Sutton
viewtopic.php?t=84525

Projecting from computer chess dev to current AI dev
viewtopic.php?t=84483

--
Srdja

[Kappatoo]

can't find you're real name, but give me a pm
and i see what i can do (if i can read the chapter)

Thanks! I just sent you a PM.

[Kappatoo]

Maybe two related threads in context "what we have learned":

"The Bitter Lesson" by Richard Sutton
viewtopic.php?t=84525

Projecting from computer chess dev to current AI dev
viewtopic.php?t=84483

Thanks for the links!
I am sympathetic to Sutton's sentiment. One thing that isn't entirely clear to me is whether leveraging computation really has to be in contrast to using human knowledge. For instance, search in chess (and elsewhere) obviously leverages computation, but it is based on algorithms developed by humans.

About the other thread:
That was interesting as well! In the chapter, I talk about the relation between evaluation and search in chess engines, including potential tradeoffs.
One thing I'm curious about (but don't really talk about in the chapter) is how well an engine could play without any search, and how neural network (i.e., evaluation) strength scales with size. There have been some recent papers by Google and by the Stockfish team related to this.

[smatovic]

[...]
One thing I'm curious about (but don't really talk about in the chapter) is how well an engine could play without any search, and how neural network (i.e., evaluation) strength scales with size. There have been some recent papers by Google and by the Stockfish team related to this.

Interesting, can you share the links to the papers?

Thomas Jahn posted his results with increasing the hidden layer for his engine Leorik, I presume with search activated:

viewtopic.php?p=978532#p978532

--
Srdja

[Kappatoo]

Cool, thanks!

I actually got confused, it wasn't Stockfish but Leela.
Links:
https://arxiv.org/abs/2402.04494
https://arxiv.org/abs/2409.12272

[smatovic]

Thanks.

--
Srdja

[Kappatoo]

Just a quick summary: The Google (DeepMind) paper claims that their neural network achieves grandmaster-level chess without search. But this is really grandmaster-level in blitz (Elo 2895 on lichess).
Then the Leela people published results showing that their neural net is stronger than DeepMind's.
I discussed the strength of these nets with Larry Kaufman. We (mostly him) came to the conclusion that Leela's net is a) IM strength in classical and b) very likely significantly stronger than any human's "intuition".
I remember being shocked by the fact that AlphaGo Zero's neural net, without search, plays at the level of a professional human player. Back then, I didn't think that was possible in chess. I've changed my mind about that. The real question is, how big would such a neural net have to be?

[smatovic]

[...]
I've changed my mind about that. The real question is, how big would such a neural net have to be?

Hmm, there are different net sizes out there for Lc0, somebody could do the testing with a depth 1 search.

--
Srdja