lightvector wrote:
Anyways, as for how to make the search capable of always solving things like this on its own... hmm...
I notice that after two steps of the ladder, KataGo (1.3.2 most recent published 20b network) starts looking at H3 as a ladder breaker (policy value 0.2%, but the "high policy" moves at P7 or P8 don't work out so well, so KataGo goes down the list of alternatives). Only after about 15,000 playouts does it start to examine G2 (policy value less than 0.1%) -- and then correctly assess how good the move is. With those numbers, of course it doesn't easily see G2 in advance.
I'm thinking this is a blind spot in KataGo's sense of shape. With KataGo training so much more efficiently than LZ, it hasn't seen quite so many million go positions, so it makes sense to me that it will sometimes miss vital points of shapes. Then again, LZ doesn't immediately get this one either...
lightvector wrote:
How does this compare to LZ?
For LZ-258, about the same. LZ wants to play the ladder, expecting black not to pull out of atari but to reply at O12 instead. When black does pull out of atari, LZ's first choice is to give up on the ladder and connect at Q7. After about 10,000 playouts, it finds G2 two moves down the search tree. The difference is that it doesn't seem to explore H3 at all, it goes to G2 first. But on medium numbers of playouts, it's choosing the same game moves KataGo, just with slightly different "reasoning" behind it.
It would be fun, but time-consuming, to compare this with a bunch of other LZ networks.
in this variation is a serious mistake as Uberdude points out.
