Miraculous Search

[michito744]

The following figure shows a wonderful search path that could never have been reached in the previous version. (Although in the actual game move-3 is D6 and move-4 is D8, it does not directly affect the path of move-5 to move-32.) The move-12 and move-13 played in the actual game were completely "beyond AI" moves. These moves were difficult for KataGo to detect even when they were a phase at pre 1 move (due to the extremely garbage evaluation value of the possible moves ahead).

Once this path was taken, the best selection would be a 1-move approach ko in White's favor, as shown in the figure. If Black can predict this far in the current situation, the best choice is to assume exchange of a ko, and C6 at move-3.

It is a miraculous achievement that KataGo1.13 reached the above conclusion at this point in the game!

[lightvector]

Cool, thanks for the report! Would you be able also to share the SGF? I would be interested in taking a look at it too.

[michito744]

This is the game record. KataGo_sample_20230604_0001.txt

Playing D6 and being received by D8 is a very bad move when competing for a 1-move approach ko.

At the time of playing move 82, Yamashita 9P did not expect this move at all, and he judged that he was in a very dangerous situation and went into a long consideration. At the move 83, Chang 9P did not expect this move at all and seemed to be upset. As a result, White retreated and Black extended its life.

Even the two players who were once at the top of the Japanese Go world could not fully grasp this phase, and the AI that evaluated during the live streaming broke down without understanding anything. If KataGo can correctly grasp this kind of phase with less energy, we will be able to evaluate the game at a different level than before.