Coffee break game: redrose vs tartrate

[xela]

A bit of frivolity for the Christmas season. People seemed to enjoy the last instalment of wild fighting, so here's a modern game that's just as chaotic.

For anyone too young to remember: tartrate appeared on KGS in 2003 but is still anonymous to this day, and redrose is the KGS username of Svetlana Shikshina. The game starts with something like a windmill only bigger. I haven't added comments of my own because I think it speaks for itself.

[bayu]

I remember! Thanks for bringing back some childhood memories.

And I was secretly hoping that someone analyses this with a bot (the same with Lee Sedol's ladder game).

[mhlepore]

Wasn't tartrate revealed to be Kim Myungwan?

[Uberdude]

And I was secretly hoping that someone analyses this with a bot (the same with Lee Sedol's ladder game).

On a few k playouts KataGo thinks playing out the not-working ladder is terrible for a while...

LSD ladder1.png (1.56 MiB) Viewed 4896 times

until it suddenly realises it's brilliant half way through:

LSD ladder2.PNG (1.58 MiB) Viewed 4896 times

I'll leave it pondering on the ladder a few moves earlier to see if it works out for itself that it's good for black.

Intersetingly LZ 254, which is usually worse at ladders than KataGo, wouldn't ladder at 66 nor 68 (would f14 extend), but as soon as black plays f13 atari it doesn't want to run out as white with only a few hundred playouts and if you do run out it ladders for black and plays Lee's tesuji so does seem to spot that at the base of the ladder. But then even stranger isntead of run out it wants white to play p17 ko threat, expects black to ignore to capture ladder and white breaks through, but then if I make black answer p17 at p16 (because white running out ladder is not a good move black needs to prevent) it does then want white to run out and thinks that's good for white and doesn't want to ladder as black! But then if you play ladder atari it again realises it's good for black fairly quickly.

[lightvector]

Getting "zeroish" bots to handle ladders properly is very, very hard. KataGo "knows" the ladder doesn't work so it doesn't like to search black to explore it much. So many black playouts leak out on every step along the way, exponentially causing black to not read the line. The cost for handling ladders well in the normal case is that you make rare cases like these much harder.

I have a (very non-"zero") hack that costs maybe 10 Elo in selfplay but that makes it very close to handling this and similar situations correctly, while still not giving up on the fast understanding of ladders in the "normal" case (where a ladder works or doesn't and there's no trick for running it all the way out anyways). Needs a bit more experimentation though. A cost of 10 Elo might be a fair price to pay for better analysis if you care about things like this.