A question & potencial improvement about MCTS

[hhfhhfhfh1]

In general, MCTS would choose the move according to the probability given by the net and the winrate of each move

but in this position, the black winrate has dropped to 0.0%, which means any other move would be better than the first choice, but it only spare a little visits(~1%) to other move which would have higher probility of winning

which was indeed the case, when remove all the analyze and analyze from the position again, the move G8 can make a draw while A8 would lead to lose, I wonder if katago can change the algorithm of MCTS: when the winrate is very low, give other move higher policy to choose, which may increase the strength of analyze.

p.s.:the net was latest kata1-b28c512nbt-s12163512064-d5648012427 and the katago version was 1.16.4

[lightvector]

Seems possible. I don't have the capacity to try testing something like this, but if you think you could try implementing it and testing it, it would be quite interesting. What functional form do you propose for this additional scaling?

(Note that nowadays, there's a good chance you could use something like Claude Code or Cursor to get an experimental implementation to help you get something working well enough to test, even if you don't know how to code that well yourself, as long as you have an idea of the kind of implementation you want and know enough to be able to prompt and run and guide things).