Fritz 17

[dkappe]

is your engine available for testing?

It isn't at the moment. An earlier version is available on github under the name leela_lite. You may have to go back a few commits to find a working version.

Most of the changes are to reduce and simplify the code and support different network types. I'll release it when it's done under the name "a0_lite" or something like that, as it really has nothing to do with leela.

My main point is that A) the code is simple and B) you can write a reasonably competitive engine in a language like python, since you only have to search in the thousands of nodes, not millions. So a slow move generator, etc., is OK.

[Daniel Shawul]

It is interesting that you consider Scorpio to be a derivative of Lc0. I strongly suspect that Daniel Shawul would take exception to this description.

He is not alone in his thoughts.

Daniel or the original poster?

If the latter, then you’ve produced a real head scratcher. Is leela chess, as the first to colonize DeepMind’s intellectual territory with c++ (there were earlier alpha zero implementations in python), the only original engine, and all others — whether derived from the alpha zero pseudo code or some other source — are henceforth clones? By that logic, all ab chess engines are just clones of Shannon’s original endgame solver. Whether that’s ridiculous is left as an exercise for the reader.

Any engine using in full or in part the code written by the Leela team is a derivative.

If Scorpio is not a Leela clone, what other explanation do you have for a mid-tier engine with an unreadable code base and a rampant history of crashes, bugs, and other play limiting or preventing issues, suddenly shooting up almost overnight to become one of the most powerful engines featured in TCEC / CCC. I won't make you guess the explanation -- the explanation is that anyone can piggy back off the work of the Leela team and with minimal effort build an engine that is a couple hundred elo weaker than Leela.

This coming from someone who has been riding the stockifsh wave for years taking its working ideas and tuning stuff...what a looser.
No, I don't consider you a cloner, but don't get it to your head that you are super-programmer of some sort while the rest of are stupid mortals.
I have seen many prodigys like you who confuse their programming skills to actual understanding of concepts ... give yourself some 5 years
and you understand eventually expoloring new territores is far more enjoyable/difficult/impactful than "tuning to death" of existing methods, especially
when there is a stream of patches coming from stockfish testing framework, which you "cloned" and renamed Ethereal testing framework apparently.
I would rather well you know try out mcts, nn or something else not my cup of tea.

For years, you have been taunting Scorpio crashing etc when it none of your business, and now that it doesn't crash at all but
it actually crashes Ethereal you are worried. I thought this would be a big lesson for you that when you are trying new stuff
,instead of tuning to death, you are bound to introduce some bugs and it is OK...

Yes I do like to give it to you when Scorpio beats Ethereal given the extra length you go to taunt me (you still do), here is another one for you:

Code: Select all

Scorpio 3.0.4 MCTS+NN Maddex   - Ethereal 11.53 x64 1CPU          35.0 - 15.0    +22/=26/-2    70.00%

[Ozymandias]

it might be worth pointing out the NN weights is not code, it is data

It is, worth pointing out and data.

The purpose of the Lc0 project is to obtain chess knowledge trough self-play. That knowledge is called weights.

As data, it may be outside the umbrella of the licence, but I wonder for how long the result of RL, is going to remain being treated as ordinary data.

[Albert Silver]

it might be worth pointing out the NN weights is not code, it is data

It is, worth pointing out and data.

The purpose of the Lc0 project is to obtain chess knowledge trough self-play. That knowledge is called weights.

As data, it may be outside the umbrella of the licence, but I wonder for how long the result of RL, is going to remain being treated as ordinary data.

No, it is just data. It is literally just a monster text file with a string of 27.9 million numbers each separated by a space. You cannot run it, just consult it. In fact that is why a GPU is so useful. Lc0 runs the basic MCTS search and for every position it asks the GPU to read the weights file to see what it says. This is incredibly time-consuming (all things being relative) which is why the NPS is so absurdly low compared to conventional AB engines.

As to the purpose of the lc0 project, it is none of what you stated. Its purpose is to reproduce AlphaZero for the PC. They have certainly moved on to try to improve on it, but that is how it started and is its purpose. As to selfplay, my first weights that played in TCEC 14 had zero selfplay, yet oddly they worked pretty well. So selfplay is a perk, but hardly a necessity.

[smatovic]

it might be worth pointing out the NN weights is not code, it is data

It is, worth pointing out and data.

The purpose of the Lc0 project is to obtain chess knowledge trough self-play. That knowledge is called weights.

As data, it may be outside the umbrella of the licence, but I wonder for how long the result of RL, is going to remain being treated as ordinary data.

No, it is just data. It is literally just a monster text file with a string of 27.9 million numbers each separated by a space. You cannot run it, just consult it. In fact that is why a GPU is so useful. Lc0 runs the basic MCTS search and for every position it asks the GPU to read the weights file to see what it says. This is incredibly time-consuming (all things being relative) which is why the NPS is so absurdly low compared to conventional AB engines.

As to the purpose of the lc0 project, it is none of what you stated. Its purpose is to reproduce AlphaZero for the PC. They have certainly moved on to try to improve on it, but that is how it started and is its purpose. As to selfplay, my first weights that played in TCEC 14 had zero selfplay, yet oddly they worked pretty well. So selfplay is a perk, but hardly a necessity.

Hehe, I do not intend to bug, but what happens if a self-play NN file is under GPL and someone uses it to add additional training on top, still GPL?

Maybe the answer is that you can not copyright NN files, but then the Fat Fritz NN file would be free to copy and modify...

--
Srdja

[Michel]

No, it is just data. It is literally just a monster text file with a string of 27.9 million numbers each separated by a space. You cannot run it, just consult it.

This is highly debatable. A compiled program is also just data which is interpreted by another program (the CPU hardware, developed itself using a hardware description language).

It seems to me that NN's may be subject to copyright, just as compiled programs are.

[dkappe]

Hehe, I do not intend to bug, but what happens if a self-play NN file is under GPL and someone uses it to add additional training on top, still GPL?

Maybe the answer is that you can not copyright NN files, but then the Fat Fritz NN file would be free to copy and modify...

You can write any old thing in a license, but there's no guarantee it would be help up in court. There are also things that you can't copyright. Certain kinds of collection are excluded. Google if you're interested.

There are two separate questions here:

1) is a weights file copyrightable?

and

2) are weights files produced using the GPL's training software covered by the GPL?

The answer to #2 is no. That's by intent. Go see the FSF's own explanations of the GPL. If you want that to be the case, you need to use a different license.

The answer to #1 is yes, though IANAL. Chessbase may require you to agree to some additional licensing terms as well. I couldn't tell you since I'm a Linux/Mac user and can't abide Wine.

[Dann Corbit]

No, it is just data. It is literally just a monster text file with a string of 27.9 million numbers each separated by a space. You cannot run it, just consult it.

This is highly debatable. A compiled program is also just data which is interpreted by another program (the CPU hardware, developed itself using a hardware description language).

It seems to me that NN's may be subject to copyright, just as compiled programs are.

If someone can prove that program output is GPL copyright, that will without any doubt be the death knell of the GPL.

[dkappe]

This is highly debatable. A compiled program is also just data which is interpreted by another program (the CPU hardware, developed itself using a hardware description language).

It seems to me that NN's may be subject to copyright, just as compiled programs are.

What you say is correct, but misses the point. The question isn't whether weights can be copyrighted, but whether the terms of the GPL apply to them (a LICENSE, not just copyright). Take the example of gcc which produces executable code from source code. This executable code is not covered by the GPL that covers gcc. Who says so? Why the FSF in their explanations of the GPL.

It is possible to write a license that says that any produced output is also covered by that license. That license may not be be enforceable, but regardless, it's not the GPL. Go read it.

[smatovic]

Hehe, I do not intend to bug, but what happens if a self-play NN file is under GPL and someone uses it to add additional training on top, still GPL?

Maybe the answer is that you can not copyright NN files, but then the Fat Fritz NN file would be free to copy and modify...

You can write any old thing in a license, but there's no guarantee it would be help up in court. There are also things that you can't copyright. Certain kinds of collection are excluded. Google if you're interested.

There are two separate questions here:

1) is a weights file copyrightable?

and

2) are weights files produced using the GPL's training software covered by the GPL?

The answer to #2 is no. That's by intent. Go see the FSF's own explanations of the GPL. If you want that to be the case, you need to use a different license.

The answer to #1 is yes, though IANAL. Chessbase may require you to agree to some additional licensing terms as well. I couldn't tell you since I'm a Linux/Mac user and can't abide Wine.

#1 sources please

#2 you got me wrong, I did not say that the output of an GPL program is under GPL, but that a NN file could be licensed under GPL or alike.

If you state #1 is yes, then I am free to put my NN under GPL, so all dkappe nets that are based on my one are GPL too.

--
Srdja