Networks for lc0

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
giovanni
Posts: 142
Joined: Wed Jul 08, 2015 10:30 am

Networks for lc0

Post by giovanni » Sun Nov 03, 2019 10:01 am

I got fascinated, but also admitdly lost, with all the recent development on programs able to use neural networks for playing. Is there any resource that would summarize how to build those networks? Also, as I don't have a GPU on my computer, would your recommend any small network to play with on lc0 or similar software?
Thanks in advance.
Giovanni

Dann Corbit
Posts: 11160
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: Networks for lc0

Post by Dann Corbit » Sun Nov 03, 2019 6:05 pm

The LC0 site has instructions on how to participate in network building.
If you want to experiment with LC0, buy yourself an inexpensive GPU. Here is a listing by price performance:
https://www.videocardbenchmark.net/gpu_value.html
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

giovanni
Posts: 142
Joined: Wed Jul 08, 2015 10:30 am

Re: Networks for lc0

Post by giovanni » Sun Nov 03, 2019 7:17 pm

Thanks, Dann. If you don't mind, have another question. Do you think that it would be feasible to feed to lc0 games of a given player to come up with an evaluation function that somehow reproduces his/her style? I know that some programs try to imitate the personality of some players (like PlayMagnus), but I am not so sure that this is the algorithm they use.
Thanks again.

dkappe
Posts: 385
Joined: Tue Aug 21, 2018 5:52 pm
Full name: Dietrich Kappe

Re: Networks for lc0

Post by dkappe » Sun Nov 03, 2019 7:31 pm

I’ve championed the cause of running lc0 on weak hardware, like integrated gpu, cpu, or even Android. It is possible to get that style of play on very modest hardware.

The best small net (128x10 as opposed to 256x20 or larger) is LD2. It gets very decent performance on cpu as long as you aren’t playing ultra-bullet. https://cdn.discordapp.com/attachments/ ... /LD2.pb.gz

There are also a number of smaller nets distilled from ID 11258. https://github.com/dkappe/leela-chess-w ... d-Networks

There are a number of other small nets, but I’d mention Mean Girl, a tiny 32x4 net. It plays on lichess looking at 25 nodes on a raspberry pi 3. https://github.com/dkappe/leela-chess-w ... -style-net

giovanni
Posts: 142
Joined: Wed Jul 08, 2015 10:30 am

Re: Networks for lc0

Post by giovanni » Sun Nov 03, 2019 8:05 pm

Nice. Thank you.

Nay Lin Tun
Posts: 599
Joined: Mon Jan 16, 2012 5:34 am

Re: Networks for lc0

Post by Nay Lin Tun » Mon Nov 04, 2019 12:56 am

giovanni wrote:
Sun Nov 03, 2019 7:17 pm
Thanks, Dann. If you don't mind, have another question. Do you think that it would be feasible to feed to lc0 games of a given player to come up with an evaluation function that somehow reproduces his/her style? I know that some programs try to imitate the personality of some players (like PlayMagnus), but I am not so sure that this is the algorithm they use.
Thanks again.

Yes, if you have a collection of 1, 000,000 games of a player.

e.g, You want to learn NN this is a cat, you need to show a million pictures of cats.

Btw, a guy in lichess has played 0.35 million games. He might be the first human in history who played > 1 million record. 🤔

jp
Posts: 1345
Joined: Mon Apr 23, 2018 5:54 am

Re: Networks for lc0

Post by jp » Mon Nov 04, 2019 12:59 pm

Nay Lin Tun wrote:
Mon Nov 04, 2019 12:56 am
Btw, a guy in lichess has played 0.35 million games. He might be the first human in history who played > 1 million record. 🤔
Are you sure this "guy" is not a bot?

Nay Lin Tun
Posts: 599
Joined: Mon Jan 16, 2012 5:34 am

Re: Networks for lc0

Post by Nay Lin Tun » Mon Nov 04, 2019 3:16 pm

jp wrote:
Mon Nov 04, 2019 12:59 pm
Nay Lin Tun wrote:
Mon Nov 04, 2019 12:56 am
Btw, a guy in lichess has played 0.35 million games. He might be the first human in history who played > 1 million record. 🤔
Are you sure this "guy" is not a bot?
No, he is not!
https://lichess.org/@/german11
https://lichess.org/player (check active players column)

DustyMonkey
Posts: 61
Joined: Wed Feb 19, 2014 9:11 pm

Re: Networks for lc0

Post by DustyMonkey » Tue Nov 05, 2019 6:42 am

Nay Lin Tun wrote:
Mon Nov 04, 2019 12:56 am
giovanni wrote:
Sun Nov 03, 2019 7:17 pm
Thanks, Dann. If you don't mind, have another question. Do you think that it would be feasible to feed to lc0 games of a given player to come up with an evaluation function that somehow reproduces his/her style? I know that some programs try to imitate the personality of some players (like PlayMagnus), but I am not so sure that this is the algorithm they use.
Thanks again.

Yes, if you have a collection of 1, 000,000 games of a player.

e.g, You want to learn NN this is a cat, you need to show a million pictures of cats.

Btw, a guy in lichess has played 0.35 million games. He might be the first human in history who played > 1 million record. 🤔
20 years ago FICS has some blokes with over 200k games ... if you have an account there, you can try fingering some of the prolific names I remember:

theblob, obliviax, andred

Raphexon
Posts: 214
Joined: Sun Mar 17, 2019 11:00 am
Full name: Henk Drost

Re: Networks for lc0

Post by Raphexon » Tue Nov 05, 2019 8:28 am

dkappe wrote:
Sun Nov 03, 2019 7:31 pm
I’ve championed the cause of running lc0 on weak hardware, like integrated gpu, cpu, or even Android. It is possible to get that style of play on very modest hardware.

The best small net (128x10 as opposed to 256x20 or larger) is LD2. It gets very decent performance on cpu as long as you aren’t playing ultra-bullet. https://cdn.discordapp.com/attachments/ ... /LD2.pb.gz

There are also a number of smaller nets distilled from ID 11258. https://github.com/dkappe/leela-chess-w ... d-Networks

There are a number of other small nets, but I’d mention Mean Girl, a tiny 32x4 net. It plays on lichess looking at 25 nodes on a raspberry pi 3. https://github.com/dkappe/leela-chess-w ... -style-net
I love your work.
Any estimation of Mean Girl's (32x4) 25 node ELO? 2500 Rapid ELO seems hard to believe.

Post Reply