I want to start with a scenario that got me thinking about how well MCTS can perform: Let's assume there is a move that is not yet added to the search tree. It is some layers/moves too deep. But if we play this move the game is basically won. However let's also assume that all moves that could be taken instead at the given game state are very very bad. For the sake of argument let's say there are 1000 possible moves and only one of them is good (but very good) and the rest is very bad. Wouldn't MCTS fail to recognize this and not grow the search tree towards this move and also rate this subtree very badly? I know that MCTS eventually converges to minimax (and eventually it will build the whole tree if there is enough memory). Then it should know that the move is good even though there are many bad possiblities. But I guess in practice this is not something that one can rely on. Maybe someone can tell me if this is a correct evaluation on my part.
Apart from this special scenario I'd also like to know if there are other such scenarios where MCTS will perform badly (or extraordinary well).