Learning from Trajectories via Subgoal Discovery

Train a subgoal policy $\pi (s_g|s_t)$ by using imitation learning. That is, by collecting a bunch of expert data.