Looking "clowny" isn't something that RL agents learn to avoid. You see it even in Go agents. The MCTS algorithm looks at probability of winning, and if it's ~1

Looking "clowny" isn't something that RL agents learn to avoid. You see it even in Go agents. The MCTS algorithm looks at probability of winning, and if it's ~100% regardless of the next move, it'll often waste moves rather than trying to win as quickly as possible.