LCZero update

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

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Locked
David Xu
Posts: 47
Joined: Mon Oct 31, 2016 8:45 pm

Re: LCZero update

Post by David Xu » Fri Mar 16, 2018 3:28 pm

His problem with AlphaZero is that it dares to suggest that the current paradigm of chess engineering might not be the global optimum, which is an unforgivable affront to... something.

gladius
Posts: 537
Joined: Tue Dec 12, 2006 9:10 am

Re: LCZero update

Post by gladius » Fri Mar 16, 2018 3:57 pm

Uri Blass wrote:
Leo wrote:Its logical to try LCZero to see if it works. I am skeptical but not a pessimist.
I look at the games and I see a lot of stupid one ply blunder that lose material.
I do not know what they do but if after many thousands of games it plays like that then I do not believe in it.
One thing to remember - the training games are played with randomization: https://github.com/glinscott/leela-ches ... #L169-L194. Eg. even if the search picks a good move, we can cause it to randomly select another one. This is done to ensure we explore more possibilities and every now and then that random selection will be better than the search, and the network will hopefully learn something :).

So, the training games are not a good representation of the real strength of the network.

That being said, it will still blunder right now. Let's see where we are in a few months :).

CheckersGuy
Posts: 273
Joined: Wed Aug 24, 2016 7:49 pm

Re: LCZero update

Post by CheckersGuy » Fri Mar 16, 2018 4:21 pm

gladius wrote:
Uri Blass wrote:
Leo wrote:Its logical to try LCZero to see if it works. I am skeptical but not a pessimist.
I look at the games and I see a lot of stupid one ply blunder that lose material.
I do not know what they do but if after many thousands of games it plays like that then I do not believe in it.
One thing to remember - the training games are played with randomization: https://github.com/glinscott/leela-ches ... #L169-L194. Eg. even if the search picks a good move, we can cause it to randomly select another one. This is done to ensure we explore more possibilities and every now and then that random selection will be better than the search, and the network will hopefully learn something :).

So, the training games are not a good representation of the real strength of the network.

That being said, it will still blunder right now. Let's see where we are in a few months :).
I watched some of the games and at least the latest ones seem like it learned "something" :P

Another thing I wanted to ask is, why doesnt it castle ?. Can anyone point me to a game where it actually castled ? :P

Robert Pope
Posts: 498
Joined: Sat Mar 25, 2006 7:27 pm

Re: LCZero update

Post by Robert Pope » Fri Mar 16, 2018 4:25 pm

I was thinking - once LCZero is trained to a higher level, is it possible to train the weights to make draws less interesting? e.g. score draws as 0.3 instead of 0.5 to make it play more aggressively for a win?

CheckersGuy
Posts: 273
Joined: Wed Aug 24, 2016 7:49 pm

Re: LCZero update

Post by CheckersGuy » Fri Mar 16, 2018 4:35 pm

Robert Pope wrote:I was thinking - once LCZero is trained to a higher level, is it possible to train the weights to make draws less interesting? e.g. score draws as 0.3 instead of 0.5 to make it play more aggressively for a win?
If it positively effects the performance then one might try that. However, I would stick with the usual scoring system as they are many other things that need to be worked on (Especially promoting the program to get more ppl involved). Restarting the training right now, doesn't seem wise.

David Xu
Posts: 47
Joined: Mon Oct 31, 2016 8:45 pm

Re: LCZero update

Post by David Xu » Fri Mar 16, 2018 4:44 pm

It already plays very aggressively (if you go to http://lczero.org/stats you'll see that the draw rate during self-play is incredibly low), so I don't see a need to try and force more aggressiveness for now.

Vizvezdenec
Posts: 52
Joined: Fri Jan 12, 2018 12:30 am

Re: LCZero update

Post by Vizvezdenec » Fri Mar 16, 2018 4:46 pm

draw rate is low only because it plays on a level of 3-years old kid.

Uri Blass
Posts: 8506
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: LCZero update

Post by Uri Blass » Fri Mar 16, 2018 4:59 pm

Vizvezdenec wrote:draw rate is low only because it plays on a level of 3-years old kid.
Exeperience shows that draw level is not so low at the lowest human level of children and you often get stalemates.

David Xu
Posts: 47
Joined: Mon Oct 31, 2016 8:45 pm

Re: LCZero update

Post by David Xu » Fri Mar 16, 2018 5:17 pm

If you actually look at the network history (which I assume you didn't bother to do, since checking the facts seems to be beneath you), you will see that the draw rate has been decreasing as training progresses. A random mover draws against itself far more frequently than the current network does.

So no, it does not simply have a low draw rate because it plays "like a three-year-old child". Wonder what it will take before you people start to understand.

Milos
Posts: 3372
Joined: Wed Nov 25, 2009 12:47 am

Re: LCZero update

Post by Milos » Fri Mar 16, 2018 5:49 pm

David Xu wrote:So no, it does not simply have a low draw rate because it plays "like a three-year-old child". Wonder what it will take before you people start to understand.
It certainly takes much more than an arrogant and delusional troll preaching them. ;)

Locked