Opening problems for AI: Problem 31 Effect of Komi in Katago

[ez4u]

The discussion of the original topic has died down. So, I will post this as a follow-up topic.

Bill's insistence on analyzing old games as if they were played with modern komi, continues to make my teeth hurt. So naturally I always start off by analyzing his positions using komi = original game. In this case that means komi = zero. Below is the result of running Katago's 40-block overnight up to 5.3 million PO's. Blue is the press at E16, just as Elf had it. However, we can see that Katago has searched some of the alternatives deeply along the way (the 4-4 in the lower right has 767K PO's for example). Furthermore we can see that the press has an estimated score of 7.7 points while all the alternatives are 7.4 points or less. In other words, the press is above a theoretical komi of 7.5 while all the alternatives are below. Of course I have whined at Bill in the past about this sort of "tipping point" being the source of Elf's "super error" judgements. So, what to do? Let's run Katago with the komi set to 7.5 instead of zero and see what win rates it comes up with...

Bills fuseki 31 with komi 0 800.jpg (145.72 KiB) Viewed 6530 times

Another night run (actually several nights later, which is why this post is so late) produced the following after 4.7 million PO's. Now we see that Katago calculates Blue as the star point in the lower right instead of the press. The difference is a fraction of a %, so easily fits Bill's "within the margin of error" meme.


Note that both the press and tenuki have estimated scores of -0.2 (i.e. both are now below the komi of 7.5 points). Also note that the search is much more concentrated on the two plays mentioned. It would be fascinating to understand whether this is random fluctuation or the way Katago will categorize its search when it is ahead or behind. (Hint, hint, feel free to comment lightvector!)

Bills fuseki 31 with komi 7 800.jpg (132.4 KiB) Viewed 6530 times

[Bill Spight]

Bill's insistence on analyzing old games as if they were played with modern komi, continues to make my teeth hurt.

I am sorry for your dental discomfort, Dave.

But I always look forward to your posts and admire your perseverance in producing a prodigious number of playouts.

[dhu163]

I'm guessing the main factor has to do with the winrate differences getting squished (hyperbolically like tanh?) as you approach 100%.

A small difference in points counts for most when around 50%, but counts for less and less as one side is in the lead.
Perhaps if 10 points lead is 93%, 20 points lead might be 98%, etc.

Another might be that different lines are chosen. When ahead, simplify and settle. This often means saving weak groups in order to get out into the centre. After the press, both sides are still a bit vulnerable locally, so B expects to add a lot more stones in that area to the relatively healthy group. Without the press, B will need to fight in tighter circumstances, with the complication of worrying about multiple groups, perhaps both playing on the top and trying to live in the upper left corner.

I'm more interested in why it dislikes the press (relatively) and also Q3.
My working hypothesis about why the press is marginally less valuable that normal is that the left is all low; and W still threatens a big moyo at the top by attacking even after the press.
Preferring R4 to Q3 seems easier to explain. B expects W to get solid with E16 whatever they play in the lower right corner. Then a shimari in the lower left or upper right won't be so big because W is solid in the upper left. Hence B wants to choose a better direction to support the moyo. I suspect the main line is R4 E16 Q15 at which point B's stronger control of the larger right side helps defend the Q15 group, which supports expansion/invasion into W's main expansion area for their moyo, the upper side. This knock on effects are very long range and hence slight, but I think I have still gained a slight sense of AI direction of play. (though I didn't manage to predict R4>Q3). If the line is R4 E16 O4, I can't explain it. However, I can't fully understand what happens with R4 Q15 (R15 Q14 R13 )? (H16)? Q13 ?