Mean Girl 7 -- the most fun leela-style net

[Dietrich Kappe]

Mean Girl 7 -- the most fun leela-style net -- has been released.

https://github.com/dkappe/leela-chess-weights/wiki/Mean-Girl:--the-most-fun-leela-style-net

[Jon Mike]

Is this another Leela clone?  

Is it just another rip off of Lc0 like DeusX?

It uses the Lc0 engine, right?  

And all you did was train the network?

What a sham!

Okay, totally joking!

(but you probably weren't laughing)

Your insightful work is always appreciated.  Without you, the only active network trainer, the community would be at a great loss.

Thanks for the human playable networks, we love the Gyals!

Great job Dietrich and much luck to future endeavors!

:)

[Dietrich Kappe]

It is such a clone! It is the cloniest of the clones. If there was a clone war between all the leela clones, it would win. It was grown from T20 DNA in a vat full of ground up baby parts. So, no.

[Jon Mike]

Hahaha!

"cloniest of clones"

LMAO!

[ronnie millsap]

anyh way to make them lower tan 2000 elo

[Dietrich Kappe]

Maybe make it a 24x3?

[Boom]

What's the smallest NN capable of playing chess?

Em sex, 16 de ago de 2019 às 12:46, Dietrich Kappe <dka...@gmail.com> escreveu:

Maybe make it a 24x3?

--
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/b624e0d1-b0a5-4778-b9b9-44f76ca94d04%40googlegroups.com.

--

Daniel Rocha - RJ

danie...@gmail.com

[Dietrich Kappe]

There’s a 16x2 net distilled from 11258. https://github.com/dkappe/leela-chess-weights/wiki/Distilled-Networks

[Mercutio]

Thank you Dietrich for another funny lco related chess engine!

It's always nice and welcome to have some unorthodox engines to play wich.

Do you know how strong your chess clo(w)ns are roughly playing (in ELO)? 

I know, it depends on hardware and time, but just a ballpark figure?

[Dietrich Kappe]

:-)

Mean Girl plays on lichess running on a Raspberry Pi 3 looking at 8 nodes per move. It has a gambit opening book and 5 man TB. It's rating (playing exclusively against humans in rapid and classical) is a little over 2400. That's frankly shocking to me. I may have to boil it down to 24x3.

I haven't really run it against anything other than LD2 (and while the bigger nets get the occasional win, LD2 is much stronger), but you're welcome to run your own tests.

[Chris Kofinas]

This gambit book is private? Thank you again for your nets!

[Dietrich Kappe]

Nope, but it also isn’t very good. :-)

http://talkchess.com/forum3/viewtopic.php?t=36090

[Chris Kofinas]

The original Dropbox link is broken. Can you share this book for us?

[Chris Kofinas]

Ok i have gambit book. I have a question: do you play on lichess with 8 nodes limit or not? I played today many games with this limit under cute chess gui and i won most of the games. It seems that on lichess mean girl plays much better!

[Dietrich Kappe]

I’ll have to check.

[Dietrich Kappe]

Ah, the config file had two entries for nodes — one for 8 the other for 25. I must be picking up the 25.

[Chris Kofinas]

I have the feeling that mini human at lichess plays with more nodes per move. Sometimes thinking for over 2 seconds. That means she oversteps the node limit. Sorry but at my chess cute games against mean girl she blundered badly with 25 nodes limit...But at 2 minutes plus 2 seconds Mean Girl without nodes limit plays very good!

[Dietrich Kappe]

I haven’t tested this on the raspberry pi 3 since switching to the 32x4 network. Given the nps of blas on the pi, it’s not getting a lot more than 25. The slow move times are usually when it gets out of the opening book and first loads the weights. I think you’ll note it doesn’t normally spend more than a second or two, and at the end of a game it has accumulated far more time than it started with. I’m not sure how to account for your different results. I’ve cleaned out all the other weights files, so it’s definitely using Mean Girl 7.

[Dietrich Kappe]

Well, this is annoying. It used to be that with thread = 1, minibatch = 1 and smartpruning 0.0 you could exactly specify the number of nodes. That doesn’t seem to be the case anymore.

[Chris Kofinas]

Sorry to spend your time to answer my questions! I want to make a stupid question... Are you sure that you offer the right weight via github for mean girl? When i unzipped the file the weight name is evilgyal-32x4-swa-200000.pb Is this right?

[Dietrich Kappe]

The only file available via the github as meangirl-7 is meangirl-7.pb.gz

https://github.com/dkappe/leela-chess-weights/releases/tag/mean-girl-7

I’m not sure where your downloading the other one.

[Eric Bonneau]

Hi Dietrich - Chris question is fully relevant; the "other one" is just the data included in the github meangirl-7.pb.gz - assuming you unzip it before use (or just to view its content).

So .pb and .pb.gz names are apparently not consistent, which is technically not an issue.

Kind regards, Eric

[Dietrich Kappe]

It is simply a compressed file, not a zip archive, so there would be no file to extract.

[Chris Kofinas]

Ok we speak for the same file (weight). I want to say this mr Kappe: when i play Mean Girl under cutechess gui with 2 minutes plus 5 seconds per move i enjoy the game(of course i use more time for me)! Mean Girl at this time control plays fantastic humanlike chess! Thanks for your work! By the way can you train in the future weights based on otb tournament games in categories for example 1800-2000 fide elo, 2000-2200 fide elo etc?

[Dietrich Kappe]

I will see what’s possible. I’m going to release 2M games of training data soon so other folks can have a crack at training with it.

[dpedley]

I'm interested in this dataset if you're posting it somewhere.

[Dietrich Kappe]

I have been sidetracked. It’s about 25GB in size. I will chop it into pieces and upload in a few days.

[Alexander Tanseco]

With all this excitement about mean girl 7, I have to ask a probably stupid question: where is all the fun I have been missing? I have downloaded the weights file and set up an engine and matched it against Stockfish but it is playing poorly. I do not understand how to use it.

[Dietrich Kappe]

Ah, I see the problem. The idea is to make the moves yourself, rather than have Stockfish do it for you. You see, there was a time when men and women played chess against each other, or against a machine, rather than having machines play against each other.

I encourage you to try it (against Mean Girl; Stockfish is absolutely no fun in this regard). It can be quite fun.

[Alexander Tanseco]

Ah, okay. Now I understand. Thanks. Hope that with Mean Girl, I get back to playing against an engine again.

[Dietrich Kappe]

Now you’re getting it! For those of us who don’t have a handy chess club, online play is too blitz heavy, and playing ab engines is like having your email spellchecked by an axe murderer, Mean Girl is a godsend.

[jaime oviedo]

Dietrich,
Probably this is a silly idea but you could train several nets of increasing strenght to help people learn playing chess. Even children could benefit from them. For instance, starting from 1000 elo, 1200, 1400, 1600, 1800, 2000, 2100, 2200, 2300 and 2400. Would you think it would be feasible? (I am thinking nets that play weaker, but sound chess, not making stupid mistakes...)

[Dietrich Kappe]

There are a few knobs one can turn: 1) smaller nets 2) fewer nodes in search 3) weaker source games in training 4) varying mcts parameters. Lots of experiments to be done.

[jaime oviedo]

I was thinking on something user friendly. Just change net and with lc0 default parameters. Maybe some catchy names like total-beginner, novice, medium, master, grand master...
(I know it is possible to change playing strenght by changing parameters but it could be tricky for a non specialist, so the hard work would be for the net trainer that provides nets to be used with default p.)

[Dietrich Kappe]

User friendly: easy.

Opponent that is weak in a particularly useful way: hard.

[Zlatko Hulama]

In all probability, to make "weaker net", I'd use a standard small net, but have it search random between one and 20 nodes, and on top of that play with "temperature" to have it choose a bit weaker move instead the "head value" - but not "blunder the queen move"!

Even the head value beats me pretty much every time, it's just too strong strategically.

I can beat the small net playing 1 node only I find some nice tactics in the opening because after it loses the material it starts playing really bad. I still lose if I blunder the material advantage...

I'm like a 900 ELO Patzer with 1900 tactics on chess and lichess. (I might be more than 900 now, but I don't play humans much anymore and just solve tactics on a toilet break)