TalkChess.com

George Tsavdaris wrote:

Dann Corbit wrote:LCZero is the most interesting thing that happened in computer chess in the last 30 years.

No way.
Deep Blue(1997), introduction of Rybka 1.0(domination for 10 consecutive years) and maybe others.
Leela is very interesting too, as its limit is the sky hopefully.

I agree.

Leo wrote:

George Tsavdaris wrote:

Dann Corbit wrote:LCZero is the most interesting thing that happened in computer chess in the last 30 years.

No way.
Deep Blue(1997), introduction of Rybka 1.0(domination for 10 consecutive years) and maybe others.
Leela is very interesting too, as its limit is the sky hopefully.

I agree.

Kai has shown that LCZero does far better than linear strength increase with increased time. The best standard programs are sublinear.

I guess few understand utterly revolutionary that is.
This is not just a new algorithm. This is a revolution.

Leo wrote:

George Tsavdaris wrote:

Dann Corbit wrote:LCZero is the most interesting thing that happened in computer chess in the last 30 years.

No way.
Deep Blue(1997), introduction of Rybka 1.0(domination for 10 consecutive years) and maybe others.
Leela is very interesting too, as its limit is the sky hopefully.

I agree.

LOL

A mediocre engine on a dedicated chess-specific hardware supercomputer, and a Fruit on steroids (bitboards + material tables), both in need of opening book to look reasonable in that stage, compared to pictures seeing chess software on a domestic GPU? You know what, pictures are worse in Chess than in Go, but it seems it still works, and in openings (and not only) even humans sometimes see pictures. This thing is surely different, heck, even I as a patzer see it playing differently. Completely differently.

I feel that LCZero is waste of time and money. Because 1) we already have 800 ELO stronger engine Stockfish and 2) even Google cannot beat SF convincingly with it's resources. Still I hope to be wrong . BTW when playing LCZero at http://play.lczero.org/ it always moves similarly.

It is certainly fun to watch its play.
Coming from a traditional alpha-beta search background and human crafted leaf evaluation function, I was astonished how well it did against IM lovik (lovlas) and GM Tang. Both times I expected it to be slaughtered, but it actually dominated its opponents in both matches. Given it was a very early NN (4.2M games) made it even more surprising - especially with the poor time management and promotion bug (although the effect of this was not too pronounced until later).
It also seems to play reasonable openings, which in itself it fascinating.

As I said elsewhere, it is good to stand-back and realise what has been done: a program has taught itself to play chess at GMish level in 4M games of self-play. And it is a _lot_ stronger now.

Whether it reaches SF level, who knows, but it is the most fascinating chess program and project I have seen for a looong time.

peter wrote:

JohnW wrote:Yeah can we get back to the Stockfish posts, we just don't have enough of them..

That's the second one secret about the A0- LC0- Hype at computerchess- fora:
Some got bored of (computer)chess as it was and now are happy thinking there was something "really" new.

As for the game (chess) itself, I fear, it won't stay for long as new as some like to think right now.
There won't be any reinvention of the "human opening theory" so much more than there was such already by engine- and database- support till now and will be further on anyhow, or it won't be "human" opening theory at all anymore.

What will a LC0- like opening book be good for humans, if they don't understand the moves and plans at least into midgame? They won't even be able to memorize the lines not understanding them.
Or do some here really think, LC0 will prove White wins with perfect play for sure from starting position?

That would be the only really new thing about opening theory to me at all, thinking till now, perfect play from both sides ends up remis most probably.

it would be interesting to see ... if some slick mathematician could come up with ... that for every move by white ... there is an equal and opposite move by black ... although ... in the quantum world .. .equal and opposite don't apply ... how close is there a connection between chess and the quantum world ...

Dann Corbit wrote:

Leo wrote:

George Tsavdaris wrote:

Dann Corbit wrote:LCZero is the most interesting thing that happened in computer chess in the last 30 years.

No way.
Deep Blue(1997), introduction of Rybka 1.0(domination for 10 consecutive years) and maybe others.
Leela is very interesting too, as its limit is the sky hopefully.

I agree.

Kai has shown that LCZero does far better than linear strength increase with increased time. The best standard programs are sublinear.

I guess few understand utterly revolutionary that is.
This is not just a new algorithm. This is a revolution.

I actually agree, and can illustrate with this utterly bizarre factoid. Bizarre for anyone who knows chess engines. I ran a basic testset with Leela and Firefly (rated 2250 on CCRL) with 100 positions. Incredibly EASY positions. Here is a sample:

[D]7r/1p2k3/2bpp3/p3np2/P1PR4/2N2PP1/1P4K1/3B4 b - - 0 1

Move is ...Bxf3+

A set that I would expect myself to solve all 100 with relative ease.

Firefly found 94/100. Leela scored 41/100. They had a full 60 seconds to find the solutions, and yes, I have a good GPU set up for it (2knps on start position). Based on any and all past understanding of engines and the importance of tactics, a match should be a massacre in favor of Firefly.

I ran my Silver Opening Suite, and stopped after 10 games due to the lopsided score of 10-0 in favor of.... Leela??!!?!?!?

I triple checked the settings and task manager, Firefly was not underusing anything.

I then set up a match against Delphil 3.2 (2494 CCRL), same conditions, 10-minute games. 20 games and will stop it too because Leela is crushing it 14.5-5.5.

I have seen engines compensate some tactical deficiencies with better positional play, but usually along the lines of 30-50 Elo at the most. Nothing on this scale. This is a new paradigm.

Mind you, its tactics are a bit of a mixed bag. In one game I saw this little gem:

[D]1q5k/8/Rprpr2p/pN2p1p1/P1P1P3/1P2B2P/3R1P1K/5B2 w - - 0 38

38. c5!! was played after 7 seconds (it no doubt saw it earlier).

I wonder if the history planes could make a difference? When you feed Leela single positions the history planes are empty whereas in a game they would not be of course.

It would be interesting to see a match between Leela and a version of Leela that has its history planes disabled...

pilgrimdan wrote: it would be interesting to see ... if some slick mathematician could come up with ... that for every move by white ... there is an equal and opposite move by black ... although ... in the quantum world .. .equal and opposite don't apply ... how close is there a connection between chess and the quantum world ...

Program is, you don't need the slick mathematician anymore since you can ask AI.

It's just as easy as that: the answer is 42!

Michel wrote:I wonder if the history planes could make a difference? When you feed Leela single positions the history planes are empty whereas in a game they would not be of course.

It would be interesting to see a match between Leela and a version of Leela that has its history planes disabled...

It would be interesting but will probably not happen very soon at least. We changed the network input/output planes twice now, and both times things went wrong, so the enthusiasm for those kind of changes cooled down a bit

When the project is a bit more stable with less rapid code changes, I'm sure it will be tried.