A comparison of the trained neural network (left) mimicking the full FSM algorithm (right). The neural network is trained on 10 million training examples from the FSM algorithm over a variety of initial conditions with standard parameter values as given above and under the ballistic assumption for the future trajectories of other agents.