On the Value of Interaction and Function Approximation in Imitation Learning