I've invented a deterministic, perfect-information game with a fairly large branching factor (~150) which tapers out dramatically after the midgame (~30 at worst). I need a strong AI. My understanding from questions like this and this is that it makes more sense to go with MCTS with a large branching factor, but as the game reaches endgame, I'm worried about MCTS missing non-obvious wins due to underestimating their ancestor game states.

I do have some ideas for a heuristic to use with minimax, but since this is a brand new game, it's probably not going to be as good as it would need to be to make minimax worthwhile in general. Instead, I was thinking about using MCTS in the early game, and switching to minimax (with alpha beta pruning) after the branching factor goes below a certain threshold for long enough.

Are there any issues with using an approach like that?



You must log in to answer this question.