From the “Computer-go” mailing list:



I wonder what “and so on” means … and I didn't know that Leela Zero was Belgian, I’d have thought she were Italian.

(h/t Ingo Althöfer)

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

I wonder if Elf means the converted weights on the Leela Zero engine or the proper Elf engine; the latter is likely a fair bit stronger.

The usual suspects have been playing on the usual servers, no ?

Uhm, I must confess that I’m not really informed (guess you mean the Fox server? Never been there), I just passed this on

But on FB “Huỳnh Nhật Tân” answered:

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

Hi Tom, yea, Fox, Tygem and other Chinese or Korean servers.

Some results and records:

== Round 1 ==

FineArt beat Golaxy

LeelaZero beat AQ

Elf OpenGo beat BADUKi

== Round 2 ==

Golaxy beat DolBaram


Fine Art beat LeelaZero



== Round 3 ==

Golaxy beat AQ
LeelaZero beat Octopus
FineArt beat Elf OpenGo in under 100 moves


== Round 4 ==
Elf OpenGo beat Golaxy

FineArt beat Octopus
LeelaZero beat Northern Light
Raynz beat Aya
AQ beat DolBaram? (or Baduki, twas Korean)

== Round 5 ==

LeelaZero beat Elf! Apparently Elf is the version Fox server have been running online.


Golaxy beat Baduki.

== Round 6 ==

LeelaZero beat Golaxy!


FineArt beat Raynz


Elf beat Octopus

== Round 7 ==

FineArt beat Aurora


Golaxy beat Raynz


LeelaZero beat Baduki


AQ beat Elf. Unexpected, big semeai at the end.


DolBaram beat Aya

Apparently the Leela Zero net is an unofficial 40 block (compared to the normal 15 block).

Final results of the prelims (top 8 progress):

1st: FineArt, 7 wins (out of 7).
2nd: Leela Zero, 6 wins
3rd: Elf Open Go, 4 wins
4th: Golaxy, 4 wins
5th: AQ, 4 wins
6th: Octopus (Leela Master), 4 wins
7th: Aurora, 3 wins
8th: BADUKi, 3 wins
9th: DolBaram, 3 wins
10th: Raynz, 2 wins
11th: Aya, 2 wins

Amazing performance by Leela Zero in this one. I went through its games with Lizzie and Leela Zero #150 weights, and wondered how there were quite many fully unexpected moves from Leela, and it seemed Leela was able to turn the tables in several of the games that LZ150 thought quite surely lost. So the 40b network trained from scratch *if I understood the Github thread correctly) using Leela Zero games (with some ELF thrown in the mix) is impressively strong.

I'm rooting for Leela Zero to get into final 4 and let's hope the semifinal is not against Fineart!

The 2nd stage of whittling 8 bots down to 4 is currently underway on Fox go server. It's all-play-all, with a game as black and white each. There's some details on https://github.com/gcp/leela-zero/issues/1554. Here are a few games.

LZ lost to Elf, reversal of result in prelims. Very strange ending, LZ played a duff ko threat that was 1 point left of the correct ko threat. I believe the game is played with gtp rather than manual relay so a misclick shouldn't be possible.


LZ beat Baduki


LZ lost both games to FineArt, here's one


Golaxy, which did perhaps surprisingly badly in the prelims given how much it was trumpeted before, beat FineArt in a weirdly short game where FA seemed to have ladder delusions, which I wasn't aware was a weakness of FA.

So according to my machine translation of this chinese website

Semifinals:
FineArt beat AQ 3-0
Golaxy beat Elf OpenGo 3-2

Final:
FineArt beat Golaxy 7-0.

So AQ ended up above LeelaZero in the 8 player round to decide top 4. There were unfortunately some misclicks (not using automatic move relay).

Here's LZ losing to AQ (from https://github.com/gcp/leela-zero/issues/1554), AQ author said he tripled computer power vs LZ:


LZ win:


A Golaxy vs Elf game in which Elf didn't realize it was dead as Golaxy's surrounding group was alive be double ko seki, so a human adjudicator ended it!

All 7 games of the FineArt vs Golaxy final are in go4go at http://www.go4go.net/go/games/twoplayer/1819/1992, I liked this example of ko threat creation and amplification from FineArt around move 75, I had previously wondered if dumb bots without abstract reasoning (as I presume FineArt is) can discover such concepts (or at least play moves which we can understand with these concepts, whether some bit of the network corresponds to this is a much harder question), the answer is yes.