How will test40 differ from test30?

Alexey Eromenko

unread,

Jan 7, 2019, 7:38:36 AM1/7/19

to LCZero

After all both are 20x256 networks.
Will t40 be trained from scratch (from ELO of 0), or forking some t30 files ?

John D

unread,

Jan 7, 2019, 1:54:18 PM1/7/19

to LCZero

Test30 was very much experimental; major parameters were changed or introduced over the course of its training and it was an afterthought for the larger community until t20 failed.

I don’t know for certain that t40 will start from scratch, but I can guarantee there won’t be any part of t30 in it.

Dietrich Kappe

unread,

Jan 7, 2019, 1:58:08 PM1/7/19

to LCZero

The network architecture of t40 will be substantially different — se and new policy convolution layer.

Jeff Wads

unread,

Jan 7, 2019, 5:21:14 PM1/7/19

to LCZero

Right from the horse's mouth.

explaination.jpg

Margus Riimaa

unread,

Jan 8, 2019, 5:26:26 AM1/8/19

to LCZero

On what basis is t40 expected to become much stronger than t30?

LuckyDay

unread,

Jan 8, 2019, 6:39:08 AM1/8/19

to LCZero

testruns done using ccrl data with se nets were >100 elo stronger than non-se nets, so that is expected to provide a elo boost.

t40 will also be trained from the start using most of the alphazero parameters, unlike t30 which has had so many changes throughout its training history that it is somewhat remarkable it has become as strong as it is, and only switched to the alphazero parameters halfway through.

so the optimism is a combination of "se nets are good" + "this will be the first proper run without bugs/huge experimental changes midway" but of course the 2nd statement may not last....

Mustafa Ünal

unread,

Jan 8, 2019, 6:40:05 AM1/8/19

to LCZero

The SE algorithm has been more efficient up to 25% in image processing. We also saw a similar performance on the T35. For these reasons I think the T40 version will be very powerful.

8 Ocak 2019 Salı 13:26:26 UTC+3 tarihinde Margus Riimaa yazdı:

Reply all

Reply to author

Forward