Game-theoretic model predictive control: Design and computational methods"

Some control problems, which emerge, for example, in vehicle traffic routing, autonomous driving, and market clearing, are characterised by interactions between multiple autonomous agents with coupled dynamics and non-aligned control objectives. Game-theoretic model predictive control addresses such systems by determining, at each time step, the control action as the solution to a finite dynamic game, namely, a Nash equilibrium. This approach enables the development of interaction-aware autonomous systems. 
In this talk, I will show a design method of the agents' objectives that guarantees stability and agent-wise infinite-horizon optimality of the nominal control action. I will then introduce a computational method based on the offline algorithmic derivation of the entire state-to-control map, and a novel active-set acceleration method for the Douglas-Rachford fixed-point algorithm. Both methods enable high control sampling times and improve the state-of-the-art computation performance by several orders of magnitude.

Join at imt.lu/aula1