Comment by Davidzheng
2 hours ago
In odds chess bots, the bots would willingly take more disadvantageous positions which are more complicated--probably the bots in GO which are trained for odds do similar? Why does it not avoid such a joseki & play a worse response which it believes the human cannot read?
Go AIs tend to naturally be quite bad at playing handicap games, due to the horizon effect. To massively simplify, when the AI sees that there's a large score gap, it realizes that every move it plays has a very low/high win rate, so it basically picks one at random. The early AIs played lots of slack moves when they were ahead, often making small endgame mistakes but winning by half a point in the end.
To account for this, KataGo uses a "playout doubling factor". When the AI plays against itself to learn, the developers set one instance of the AI to use fewer playouts compared to the other one, but gives the weaker AI some handicap. This allows the AI with more playouts to learn that although it may be in a losing position, if it makes the board position chaotic enough, it may still win.
The flying knife is objectively an extremely complicated position, so the AI played it assuming that the opponent would be forced into a very complicated reading battle where they could make some mistakes. Unfortunately, Shin has memorized the flying knife joseki more thoroughly than any other human on the planet, so he could play exactly like a very strong AI. It would probably be possible to train an adversarial network specifically to beat players like Shin, but that would take a substantial amount of effort, and Shin is strong enough that it probably wouldn't make too much of a difference -- Shin won by 11.5 points in game 3 without a flying knife shenanigans, only losing 7 points of value throughout the entire game.