TCEC season 13, 2 NN engines will be participating, Leela and Deus X

[Uri Blass]

Well, there is extreme high possibility that Deus X authour is Deep Junior Authour, Shay Bushinsky.
https://www.aaai.org/ojs/index.php/aima ... /view/2255

So you will be seeing updated version(work of art) of Junior!!

No - you're wrong, but my lips are sealed.

What about Tencent company?
https://www.chess.com/forum/view/genera ... -season-13

http://www.chessdom.com/deus-x-the-nn-c ... rt-silver/

My God, what a load of BS. I would think that submitting main engine and then it's tweaked clone is something that wouldn't be allowed by TCEC rules, but apparently it it going to be.

For once I have to agree, that's pathetic and ridiculous at the same time.
But that was already clear from the first cryptic announcements though.
The most work done was to find a new name...

ASilverheute um 06:08 Uhr
It is a 100% unique NN built entirely from human games, and only human games, from Megabase 2018. No self-play, or anything else. It is 10x128 for any curious.
That was the philosophy behind it.

Armstrongheute um 06:09 Uhr
Sorry for my naive question . Not gone through the video . If Deus in agreement with the open source policies of Leela ?

ASilverheute um 06:09 Uhr
What open source policies? I am not selling it.

Now you can start with dozens of Brainfish clones (with different self constructed 'cerebellum' books) and dozens of LeelaChess clones with
self-trained NNs. Brave new world.

I think that there is a difference because I guess that Brainfish derivatives are similiar in the choice of moves when LeelaChess derivatives with a new net are not similiar in the choice of moves but I am not sure about it and it may be interesting to see the difference in move choice between LC0 and Deus and compare.

[Branko Radovanovic]

I think that there is a difference because I guess that Brainfish derivatives are similiar in the choice of moves when LeelaChess derivatives with a new net are not similiar in the choice of moves but I am not sure about it and it may be interesting to see the difference in move choice between LC0 and Deus and compare.

Agree, this is not comparable at all. Different net - trained in a completely different way -> different engine.

Also, let's not forget there is still no evidence the "zero approach" (i.e. not relying on any human knowledge) is optimal.

If I had to create a NN engine which improves upon Leela, I'd take the hybrid approach: compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net. This could result in better eval, enable one to reduce the net size (and thus improve performance), or both. If I wanted to do this as an open source project, I'd certainly use the existing Leela code base rather than develop everything from scratch. Still, functionally this could not be called a "derivative".

NN engines that use CPU calculation of net inputs might be in the works already and I'm fairly certain that we'll see them soon. In the context of TCEC, that's going to raise the issue of fairness with respect to hardware use. I definitely don't think dual 1080 Ti against 43 cores is unfair, but dual 1080 Ti and 43 cores against 43 cores alone is unfair, and everything in between is going to be difficult to judge.

[Milos]

I think that there is a difference because I guess that Brainfish derivatives are similiar in the choice of moves when LeelaChess derivatives with a new net are not similiar in the choice of moves but I am not sure about it and it may be interesting to see the difference in move choice between LC0 and Deus and compare.

What the hell are you talking about???
Do you know how many lines of code of LC0 Albert changed? My guess is less than what a typical cloner of SF does.
He has identical net, identical number of inputs, levels, filters, using exactly the same backend, same search.
What he did is identical to taking SF, changing PST numbers and evaluation bonuses (like what Tsvetkov was doing) and claiming the new engine that is btw. probably hundreds of Elo weaker than original SF. You really think that changing only some numbers randomly in SF eval wouldn't make it play totally different moves?
And after all there is question of legality, since TCEC is public event and that "Deus X" competes on it, if he doesn't publish the source code he is in direct violation of not only GPL licence but also NVIDIA licence.
If you really want to see innovative NN engine look at Scorpio MCTS+NN of Daniel. Totally different net, backend, search, everything.

[Milos]

If I had to create a NN engine which improves upon Leela, I'd take the hybrid approach: compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net. This could result in better eval, enable one to reduce the net size (and thus improve performance), or both. If I wanted to do this as an open source project, I'd certainly use the existing Leela code base rather than develop everything from scratch. Still, functionally this could not be called a "derivative".

What you write there is equivalent to:
"If I wanted to cure cancer I would do the following..." or "If I wanted to build a rocket to fly to Mars, I would do..."
I mean equally efficient.

[glennsamuel32]

From the interview Albert says (quote)
"As to whether it will be commercial, who can say...
If I come out with something that's really spectacular and that outdoes Alphazero, very possibly.
Right now at the moment, it's a private project."

Since February, the gpu's are crowd-sourced.
From all the knowledge attained from those 22 million games, comes private commercial ventures.
Doesn't make any sense, especially to those who run their hardware 24/7, thinking that they're contributing to a public venture.

[Branko Radovanovic]

If I had to create a NN engine which improves upon Leela, I'd take the hybrid approach: compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net. This could result in better eval, enable one to reduce the net size (and thus improve performance), or both. If I wanted to do this as an open source project, I'd certainly use the existing Leela code base rather than develop everything from scratch. Still, functionally this could not be called a "derivative".

What you write there is equivalent to:
"If I wanted to cure cancer I would do the following..." or "If I wanted to build a rocket to fly to Mars, I would do..."
I mean equally efficient.

Not at all, this is a very simple and also rather obvious idea.

[Milos]

If I had to create a NN engine which improves upon Leela, I'd take the hybrid approach: compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net. This could result in better eval, enable one to reduce the net size (and thus improve performance), or both. If I wanted to do this as an open source project, I'd certainly use the existing Leela code base rather than develop everything from scratch. Still, functionally this could not be called a "derivative".

What you write there is equivalent to:
"If I wanted to cure cancer I would do the following..." or "If I wanted to build a rocket to fly to Mars, I would do..."
I mean equally efficient.

Not at all, this is a very simple and also rather obvious idea.

It is such a simple idea that since LC0 appeared no one actually even tried to change the format of NN let alone "compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net". I am sorry, but I don't think you know what you are talking about.

[Branko Radovanovic]

It is such a simple idea that since LC0 appeared no one actually even tried to change the format of NN let alone "compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net". I am sorry, but I don't think you know what you are talking about.

What makes you think that?

[Milos]

It is such a simple idea that since LC0 appeared no one actually even tried to change the format of NN let alone "compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net". I am sorry, but I don't think you know what you are talking about.

What makes you think that?

Because things like "compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net" make no sense at all.
Output of computations of conventional chess evaluation terms is nothing but numbers in centi-pawns. NN takes board representation in some format and extracts features. These two are worlds apart. Combining them is pointless and very difficult if not impossible.

[Gian-Carlo Pascutto]

It is such a simple idea that since LC0 appeared no one actually even tried to change the format of NN let alone "compute some chess evaluation terms in the conventional way, and provide them as additional inputs to the net".

The people involved in the Leela Zero Chess project obviously like to keep the "Zero" aspect and not introduce bias by pre-selecting evaluation terms. That doesn't mean it's not a very obvious enhancement for anyone else to use that doesn't feel bound by such a restriction and wants to train some nets on their own.

That being said, such an approach is likely only useful if you are working on an engine containing a small/fast network. For a pure NN engine, I doubt the benefit of saving a few layers is a significant factor in overall performance. But it's still a "free" gain if you don't care about "zero".