Reinforcement learning was applied after the basic model was initialized with imitation. Maybe that can partially explain the small number of steps.