lightvector wrote:
If you set the komi to -2.5, KataGo thinks the game is much closer to fair. It thinks black is still 0.5 points behind, which is not entirely consistent with its estimation of black in the original state being behind by 0.5, since black has just gained 10 points over its initial estimate of black behind by 9. Possibly this is due to the search now exploring significantly different lines *after* the corner sequence due to the closer winrate for how black should handle the white center group.
In the search for the margin of error for winrates, I considered using the log of the odds ratio instead. However, I decided not to bother, because it seems to me that when the winrate is near 0 or 100%, the search will be greatly affected. So I just take winrate estimates less than 10% or greater than 90% with a large grain of salt. I don't know if this matters for KataGo and point estimates, however.
I like the idea of setting komi with KataGo for analytical purposes. For one thing, my guess is that the top choices will be closer to correct play, instead of desperation measures. That's closer to how humans try to do analysis, anyway. As for making plays that put the game on the line (勝負手) or looking for a place to resign, we humans are rather good at that, when playing against other humans. The bot equivalent often strikes us as silly.
Quote:
It also thinks that one of white's moves on the corner saving variation is a significant mistake, white can do better if black tries to play this way. Given that at this point, KataGo should usually be quite a bit stronger than ELF, and that at least its preferences line up pretty well between the komi-adjusted and non-komi-adjusted cases, I'm inclined to distrust ELF here about what the "correct" variation is, except insofar as maybe ELF could be evaluating the winrate correctly "for itself" as Bill suggested.
There is a phenomenon with chess engines that may appear with go bots, as well, that humans are not able in some positions, to apply the AI evaluations. GM Ben Finegold sometimes points out that engines may evaluate the position his is discussing as even or nearly so, but that he and other human GMs much prefer one side or the other, typically the side with an attack, because the engines know how to defend against the attack, but humans don't.
Login
Register
FAQ
Search
lightvector
should be the kick at a, and also that 
is better than the atari at b.
, 
makes the solid connection, and then 
pincers the two Black stones. White is playing catch-up. Perhaps we can consider this pincer a shobute.