Learning Representative Trajectories of Dynamical Systems via Domain-Adaptive Imitation

This lead to DATI, a robust method to learn representative trajectories of dynamical systems by framing the trajectory imitation task as reinforced style transfer problem from the reference trajectories to the rollouts of a reinforcement learning agent.

The method can be used for spatially-unconstrained trajectory data mining. In particular, we applied it to study anomaly detection from tracking data, using maritime traffic as a testbed.

You can find a demo of a typical rollout here.