A model that generates sequences of steps (trajectories) by sampling from a learned policy at each step.