An analysis engine that helps compare different moves fairly (to the extent that is possible) using KataGo.
KataGo is a very useful tool for analyzing Go games. However, it can be hard to know how to interpret its output. When comparing moves, which metric is best to use: score lead or win rate? How many visits are needed to estimate these values accurately? This program helps address all of those questions.
Multi-threaded search is generally weaker than single-threaded search given the same number of visits.
KataGo's search behaves differently depending on who is ahead. The only way to ensure that the strongest moves are being searched is to make it believe that the game is even.
Not all visits are equal. KataGo estimates how accurate its evaluations are for each visit. That uncertainty is taken into account by using weights instead of raw visits.
Radius shows how large the variance of the utility score is. The larger the radius, the greater the variance.
Utility score is what KataGo uses during search. It includes estimates of win rate, final score, and its variance. Because the evaluation is based on an "even" position thanks to adjusted komi, this is a strong metric for comparing moves.