Learning Interactions and Dynamics of Swarms

This paper compares and analyzes several existing methods of nonlinear System Identification (SID) for the self-propelled particles swarm model. Methods that range from simpler data-driven techniques such as Sparse Identification of Nonlinear Dynamics (SINDy) to more complex learning methods such as RNNs, CNNs and Neural ODE have been explored. The aim is to predict future trajectories of the swarm by approximating the nonlinear dynamics of the swarm model. We experiment modeling with (1) transient and (2) steady state data respectively from a swarm simulation. We demonstrate that Neural ODE, combined with a carefully selected model trained on transient data is robust to different initial conditions and can predict the correct swarm stability, outperforming other learning methods.

Skills: Deep Learning, System Modelling and Identification

Report available: here