Maybe it's true that intelligence depends on the environment, but consider: the environments where policy iteration performs better than RL with temporal difference learning are kind of dumb.