Vad ca s-a testat network-ul #136 (ultimul la momentul asta) cu network-ul ELF si distanta dintre ele a ramas cam aceeasi
GN: Have you been able to make money from Leela Zero?
GCP: No. But I think it has made a lot of Go players happy.
Small note because I know many people will start "panicking" as we have 200k games with no notable strength progress (i.e. PASS).
One reason why progress dropped a bit is that the ELF games left the training window. This wasn't intentional.
How many Gigabyte of data is the necessary training data ?
A full window is about 20G-30G IIRC.
The problem is that it evolves a few times a day so it's not very fit for torrenting. The storage server that broke was very nice because I could update it almost incrementally with rsync.
At worst I'll rent a dedicated server with a big HDD for a while.
So far, it's been lowering learning rate, and increasing the network size. If we're at the lowest rate at 256x20 I personally think that's a sign we should look towards something else than 19x19 @ 7.5 komi that is more beneficial to Go players
Code: Select all
2018-06-27 15:46 62b5417b VS 12692a83 - 334 : 68 (83.08%) 402 / 400
Code: Select all
2018-07-16 01:29 62b5417b VS 050375ce - 310 : 99 (75.79%) 409 / 400
Looking to the future, at some point I think it'll be necessary to have a production and a test side, split the clients say 80/20 and on the test side, run experiments where you restart from scratch and see how much (more) progress you can make in say 1-2-3 weeks compared to the baseline. Leela Zero Chess has been doing this, and I think it's a decent approach to try to find fundamental improvements.
Yes, 256x20 is still the plan for now, as soon as we have a suitable network. It's bootstrapping the latter that 's proving a bit difficult now. If there's no good 256x20 but we have a good 256x40 I'm certainly willing to just that that too.
Users browsing this forum: No registered users and 2 guests