Lc1

[Cscuile]

Github.com/Cscuile/Lc1

Lc1 will take the non-zero approach to produce the strongest chess entity possible.

I will first mention that this is NOT a 1 man project. I will need help. The first step is to collect as many high quality games as possible from sources like CCRL. 

Please post a link into "High Quality Games Link" of the source as well as a download link to the games. 

Thank you, 

Cscuile. 

(It is about time Lc1 was started) 

[Shah]

Great, Cscuile.

But naming it LC1 and in the same breath say it is not a 1 man show... (-;
(Just kidding of course, good luck with it!)

[Carl Bicknell]

Are you Sven Steppenwolf?

Am I right in saying you won't be generating new games but training with existing ones which can be collected from anywhere possible?

[Margus Riimaa]

I love your idea.
Are there any ideas to make the NN less general and more chess specific - such that it has the structure on chess built in in its structure?

[Margus Riimaa]

*structure OF chess

[crabe1054]

In the same idea , is it possible to train a NN only on GMaster Magnus Carlson games o rGMaster Boby Fisher Games ?

Could be grea

Le dimanche 24 février 2019 11:49:29 UTC+1, Margus Riimaa a écrit :

*structure OF chess

[denochss]

Hi everyone, I think the idea for an artificial intelligence is to learn from your game and your mistakes, either play against yourself or against another engine. The Leela learning database should be about games where one of the engines that play the game is lc0.
Greetings.

[Cscuile]

I really hope I can get some help with this. Thanks!

[Cscuile]

With Lc1 we could! Anything goes as long as it produces the strongest network, and so if Lc1 gets to the point in which training against SF does not produce any spectacular results, self play is then the way to go. 

[Cscuile]

Hmm, I'm not entirely sure about NN structures that benefit chess. The only one that comes into mind are SE, however I don't think people have specifically researched structures for chess. 

[Cscuile]

Of course, if you have enough games. 

[Esteban Canepa]

About NN for chess: what about using the same net but add to the inputs attacked squares boards by piece type or something like that?

[Deep Blender]

Providing this sort of additional input planes may indeed help to stabilize Leela. A neural network has a hard time to learn that there can be a huge difference if a piece is just shifted by one position. That's likely one of the reasons why history planes are needed. However, if there are additional planes for attack positions or even movement positions by piece, this could potentially increase the training speed and potentially make history planes obsolete.

[OmenhoteppIV]

Cool idea.. supervised learning. but deepmind already did that.. the results is pretty strong but not as strong as Zero.

[NuclearPawn]

Hey Cscuile!

I Play Correspondence chess, and when it comes for quality I think CC has one of the highest, if not the highest quality. Here Attached I will give you all the World Championship Correspondence games from 1950 up to 2019 in a pgn format. I am not posting the link since these games might not be available for download unless you are ICCF member.

I have more archives of high quality CC Grandmaster's games if you're interested. 

Good luck on the project!

  

[NuclearPawn]

This is a link (116mb Zip file) of all the games on ICCF server also. I don't know if you would be able to open and download it so let me know.

http://gamesarchive.iccf.com/Complete.zip 

[Hope]

What would be interesting to me is if Leela can be combined with a traditional chess engine to improve her tactical play.

[Pawel.NewYork]

Those early years games arent such an amazing. I just checked a few of them from 50' and they dont seem to be a high quality ones from an engine perspective. You dont want to feed  NN  with low quality games so it could produce a low quality output, dont you? 

[denochss]

El domingo, 24 de febrero de 2019, 7:28:48 (UTC+1), Cscuile escribió:

 Let's distinguish Leela's learning processes. Lc0 learns making games automatically. Another variant of Leela is that of AntiFish, where the confrontation is reinforced by playing Lc0 against another engine such as Stockfish, in higher quality games. Finally we have Leelestein or Lc1 that learns from all the bases of items supplied, bases similar to CCRL. The best option to my understanding is Leela Antifish. 

Greetings.

[jessi]

I like the approach lc0 => lc1 including all kind of good games as well as selfplay games :-)

Is it possilble to train on positions (with their solutions) too? Endgame positions might be usefull.

Did you already ask Cato at TCEC for his famous catobase?

You know the archives of TCEC (esp SUFI) and CCCC very well, aren't you? Sure there aren't many games but high quality.

[Cscuile]

Nope I haven't. But if you have a link, please do add it to the github!

[Cscuile]

I should mention again, I likely will not have time to develop Lc1. I just want to see it started.

I will need the communities help from this or else it may never get started. 

[Jeremy Zucker]

Hi Cscuile,

  I am glad you started this thread.  I have been thinking about how to train lc1 to achieve its goals, as well as some of the subgoals that folks on this thread have suggested, such as lc-stockfish, lc-CCRL, lc-fischer, lc-carlsen, etc.  It seems that this is really an exercise in transfer learning. That is, rather than start from scratch (the lc0 approach), lc1 could start with the lc0 network, and then freeze all but the last few layers, and train the remaining layers to play in whichever style you would like using a (much smaller) dataset of games to emulate.  For example, one can take a subset of positions from Fischer games, train the value head on the outcome of the game, and train the policy head to prefer the actual next move played by Fischer.  The validation set would be the subset of positions that were not used for training.   When the error on the validation set drops below a certain threshold, you can say you have trained the network to play in Fischer style.  

Same goes for CCRL, stockfish, tablebases, Carlsen, Tal, or any mixture of the above.

Sincerely,

Jeremy

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/717eaa0d-69a5-448e-b89d-cefdf5413eb1%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

[jimgl...@gmail.com]

I ask because I do not know. Is it possible to...

1) Take a trained Leela network like 33000 or whatever and (a) add training on a wide range of openings that it hasn't trained itself on, plus (b) integrate Ender to take over the endgames?

BTW, what's happened to Ender?

2) Take an engine like Stockfish, keep its Alpha-Beta search for comprehensiveness, but delete its 'formula' positional evaluation function and substitute the evaluations of a neural network like Leela 'learned' from millions of self-play games?

Just asking.

[Cscuile]

Note1: One idea was to train a 40b SE network with the collected games.

[Cscuile]

Note2: Generate 10M very high quality SF vs AntiFish games?

[Andrew Palfreyman]

You will need upwards of 40 million high quality games. How do you intend to get hold of them?

[denochss]

El martes, 26 de febrero de 2019, 20:04:53 (UTC+1), Andrew Palfreyman escribió:
> You will need upwards of 40 million high quality games. How do you intend to get hold of them?

You have to ask Google.