AI Jubango

General Go Discussion
1

In this topic, LeelaZero goes head to head with Minigo in a traditional jubango. The crowdsourced AlphaGo re-creation we all know and love pitted against the squeeze-and-excite advances of rogue Google employees in a cosmic boxing match for the ages. Finally, I get to see which is stronger! Ringside seats are right here folks, the epic bout of infinite possibilities, LeelaZero and Minigo dressed in silk shorts and padded gloves, clenching mouth pieces, throwing punches! And, of course, the chance to see one or the other lose a game for once!

(Settings are 2000 visits for LeelaZero v0.17 40-block network 232, and 4210 visits for Minigo v17 19-block network 961. I’m sorry if the settings are crap. I chose the strongest networks as displayed on cloudygo.com and chose visits for equal time at a decent pace for my pc. Suggestions are welcome, bearing in mind this is for fun and totally not official…)

Game 1 (q16): LeelaZero plays the dagger variation of the 3-3 invasion (aka “new taisha”). Relentless fighting in which a corner stays undeveloped for some 160 moves.

The winner is...

Minigo!

Game 2 (q16): It’s Minigo’s double approaches versus LeelaZero’s double 3-3 invasions. Which plan succeeded this time?

The winner is...

Minigo!

Game 3 (r16): In which a complex corner variation results in inadvertently running out a bad ladder.

The winner is...

LeelaZero!

Game 4 (r16): In which AI falls for a stupidly simple kyu-level ladder and sees its win rate plummet to 20% but wins anyway.

The winner is...

Minigo!

Game 5 (q16+r4): Forget orthodox and approach openings; just drive a dagger into the 4-4 point and fight for the rest of the game.

The winner is...

Minigo!

Game 6 (q16+r4): Nothing can save AI from bad ladders (when there’s two of them within the first 100 moves).

The winner is...

LeelaZero!

Game 7 (q16+q3): The dagger fails to slay the opponent yet again.

The winner is...

Minigo!

Game 8 (q16+q3): LeelaZero shows some variety in its 3-3 point invasions, but the fighting is no less relentless.

The winner is...

LeelaZero!

Game 9 (r16+q3): Another dagger, despite the presence of two unenclosed 3-4 points, which is meaningless to LeelaZero.

The winner is...

LeelaZero!

Game 10 (r16+q3): A more classical opening pattern with enclosed corners, approach josekis, and base-making.

The winner is...

LeelaZero!

Final tally...

Minigo 5 : LeelaZero 5

Thanks for reading. :slight_smile:

13 Likes
2
Instructions unclear.
instructions-unclear-3402070.png500Ă—689

So… there’s an LZ network #132 which clocks in at 15 blocks and a selfplay ELO of ~10800.
There are also in total 61 40b networks ranging from #174 (12463) to #234 (14836).

What’s the LZ network’s hash?

1 Like
3

Ahh, it’s official network 232 (hash 06e03050).

3 Likes
4

Yea iirc minigo got in total about 30M games in before it went on hiatus. LZ is sitting at a mere 15M so far.

5

@mark5000 Thanks for sharing! How do you get minigo to run? LZ engine with the weights from your link?

Side note, I recently discovered KataGo thanks to @smurph. Currently it’s about the strength of ELF v2, I’m looking forward to their next run. That could be the next jubango challenger against the winner of this match :slight_smile:

2 Likes
6

Yes, it is in fact LZ Minigo with converted weights as found here: https://github.com/JYPark09/leela-zero/releases/tag/minigo-v17

I could run KataGo against the victor or do whatever the crowd wants next. I’m pretty certain KG is slightly weaker than both. Its creator rated it comparable to an older LZ network and equal to ELF OpenGo v2.

2 Likes
7

Thanks! And right about KataGo’s strength, that’s why I suggested waiting for its next run. Contrary to the others, KataGo gets injection of technical knowledge (like ladders) which helps it improve faster.

2 Likes
8

Games 5 and 6 are up.

9

This Jubango is over.

Congrats to...

both, because the result was a draw!

5 Likes
10

So should we conclude that some bots handle the pressure better than others during major events?

4 Likes