Actually, both KataGo and Leela Zero require that the highest LCB move also has enough playouts compared to other moves, not just 10, otherwise LCB would be a great harm on the strength instead of a benefit. It's a bad idea to pick a barely-searched move when the barely-searched move might only falsely seem good due to not having been read out enough to find the correct tactics yet. For KataGo, it's currently required by default that the LCB move has at least 15% of the playouts that the top playouts move has.spook wrote: Afaik, by default KataGo just takes the move with the highest LCB, as long as it has more than 10 playouts. Although, I wouldn't be surprised if this was configurable to some degree. Then again, even if there is, tweaking KataGo, wouldn't be a solution, because Leela Zero has the same issues.
But zbaduk probably should not hardcode any such rule on its own. The final word on the preference ordering of moves should be given by the "order" field returned by the analysis. Zbaduk should simply just be sorting by this "order" field, whose purpose is specifically to let the bot give its preference ordering according to whatever combination of value/playouts/policy or whatever internal metrics it deems best. I'd be surprised if the "order" field were ranking the 80-playout move at the top, unless KataGo has also been somehow explicitly configured to also override the 15% threshold as well, so this would fix all the issues like the one you just observed that arise from trying your own custom sorting logic.
