Life In 19x19

https://www.reddit.com/r/baduk/comments ... ship_2020/

_autoencoder_ wrote:Hey everyone!

Big news. We are going to be streaming the GO AI World Championship tonight at 9:30pm ET.

I will be hosting and Bob Song, an officially certified Chinese 5d amateur, will be joining for commentary and game insights!

The first match is LeelaZero vs. Golaxy. Here's where you can find it:

Twitch: https://www.twitch.tv/gameofgo
Facebook: https://fb.me/gameofgoapp

YouTube: https://www.youtube.com/channel/UCXqY9R ... TzeEUUAjlA

Hope to see you in the chat! Should be fun

https://www.reddit.com/r/baduk/comments ... ournament/

It's the first international tournament for KataGo! Leela is also participating, and the favorite should be Golaxy (I don't see FineArt in the list).

This competition is held in Fuzhou, China, and I am a Chinese, but I can't look the live of this competition!
First, I can't see those sites like reddit, youtube, fb and twitch because of the great wall!
Second, I can't find a live of this competition any Chinese sites!

Anyone know if the engines were optimised for the given hardware? It's not uncommon for tuning to be required to get best results.

some results here
and some comments here

3 timeout and an invalid game in the first round and 4 timeout in the 2nd round

I don't see the hardware specified any of the links. Is it uniform or each team brings their own?

Updated results, with... 10 timeouts in the last 3 rounds... HERE

From Bob Song on discord:

Final results after the qualifying round (top eight advance):
1 MyrtleGo
2 ChaoRan Go
3 Golaxy
4 TianGo
5 Yile Go
6 DolBaram
7 Leela Zero
8 Yinzhe Go
...
9 Katago eliminated

Yike says the katago operator made several mistakes resulting in losing a won game due to time, and decreasing the visits because he underestimated the opponent, which also lead to a loss.

Is there a page with background information on each program. The language it's written in, its authors, blah blah do dah

How KataGo operator could make mistake when engine is connected by GTP? Also do operator of KataGo is somehow connected to creator of it? What about LZ? LZ won 4 games, and lost 3 (all by time). Sorry, but for me all of these sounds fishy.

The KataGo developer is not involved with whoever is running "KataGo" in the tournament. (source: Katago Discord)

At this point the whole tournament seems like a pointless exercise with so many timeouts from different programs. This would seem to point to a problem with the organization of the tournament, whether lack of clear communication or coding issue. There are few enough rounds that the timeouts are not just in the noise and seem to have a large effect on which programs advance.

Even before the tournament, lightvector, Katago's developer, says the it should just be considered as entertainment, and we shouldn't take the results too seriously. Real tests of strength come on equivalent hardware and should be tested with many games.

That said, I am going to start looking at the games. As I do, I can post them here. For example, here is the final round game with katago versus golaxy. Reviewing on katago, it agrees it is way behind pretty early. Either Golaxy was reading way deeper (so better hardware) or katago has a blind spot. Cool game. Should I post games I find interesting?

How did the officials prevent the bots from cheating by conferring with humans?

Seems this tournament used a variation of button go.

According to Sensei's Library (https://senseis.xmp.net/?2020WorldArtif ... ompetition) the scoring rule is:

6. Chinese rules, komi is 3.25 stones but if white passes first it's 3.75.

This is similar to WMSG rules in 2008 (https://senseis.xmp.net/?WorldMindSportsGamesRules), which had:

If White passes first, Black's score is reduced by 1. If Black passes first, no such compensation is given.

Yes that sounded stange, I'm pretty sure LZ (and maybe many other simpler bots) cannot be aware of this.

LZ timeout in the first game against ChaoRanGo.