Mean Field Control in Urban Traffic

We propose to model traffic scenes as particle systems with pairwise interactions, leading to a distributed optimization problem that can be solved using consensus algorithms. We show that the procedure is amenable to an algorithm with linear space complexity via a measure-theoretic formulation. We propose a Lagrangian based on the Christoffel-Darboux kernel that can be estimated from data and illustrate our method on a real-world roundabout scene. Our implementation is available at https://github.com/ebuehrle/manybody.