Page 1 of 2

LC0 - how to catch up?

Posted: Mon Apr 16, 2018 9:34 am
by smatovic
Stunning progress in the last weeks,
maybe it is a bit too early,
but i believe here is an paradigm shift underway.

So how do the old school chess programmers intend to catch up with LC0?

--
Srdja

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 10:35 am
by Vinvin
Image

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 12:18 pm
by Uri Blass
smatovic wrote:Stunning progress in the last weeks,
maybe it is a bit too early,
but i believe here is an paradigm shift underway.

So how do the old school chess programmers intend to catch up with LC0?

--
Srdja
I do not see a progress in the last weeks but a regression.

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 12:33 pm
by smatovic
Image

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 12:35 pm
by smatovic
I do not see a progress in the last weeks but a regression.
I mean the last weeks since the selfplay and training started.

--
Srdja

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 12:36 pm
by Henk
I don't even know whether development using zero method or alpha beta method cost more electricity. Probably the first I guess.

They told me in 1994 if you know no better algorithm then use neural network. Don't know if that is still valid.

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 1:01 pm
by smatovic
They told me in 1994 if you know no better algorithm then use neural network. Don't know if that is still valid.
They told me in 2008 the same for Monte-Carlo methods...

--
Srdja

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 4:21 pm
by Hai
Uri Blass wrote:
smatovic wrote:Stunning progress in the last weeks,
maybe it is a bit too early,
but i believe here is an paradigm shift underway.

So how do the old school chess programmers intend to catch up with LC0?

--
Srdja
I do not see a progress in the last weeks but a regression.
That's bad for you if you can't see the progress, because I can see a big progress.

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 4:40 pm
by Ras
Three things:

1) Alpha0 had that stunning hardware optimised for NNs although frameworks for GPUs already existed - but they wouldn't cut it, that's why Google came up with TPUs.

2) Low hanging fruit. Progress becomes harder over time.

3) You can't actually debug an NN.

Re: LC0 - how to catch up?

Posted: Mon Apr 16, 2018 5:09 pm
by Uri Blass
Hai wrote:
Uri Blass wrote:
smatovic wrote:Stunning progress in the last weeks,
maybe it is a bit too early,
but i believe here is an paradigm shift underway.

So how do the old school chess programmers intend to catch up with LC0?

--
Srdja
I do not see a progress in the last weeks but a regression.
That's bad for you if you can't see the progress, because I can see a big progress.
I read in the forum and look in the graph and in the last days I do not see a progress based on the reports.

There was a progress in march but the graph shows no progress in the last days and it is not only the graph but also what people report here.

http://lczero.org/