Learning Agent-Based Model Predictive Control for Holistic Vehicle Performance
Original title:Learning Agent-based Model Predictive Control for Holistic Vehicle Performance
Distributed agent-based model predictive control frequently encounters limitations when cross-agent interactions cannot be fully modeled in advance. This paper introduces LAMPC, a hybrid architecture that incorporates online Gaussian process regression into distributed predictive control. By estimating the mean and variance of unknown agent contributions, the controller updates system predictions over the horizon while enforcing soft chance constraints. Validated via simulations and vehicle experiments, the method preserves real-time operational feasibility across flexible topologies, bridging data-driven uncertainty estimation with deterministic safety bounds.
Why it's worth reading
It provides a mathematically grounded hybrid control formulation, coupling Gaussian processes with chance-constrained MPC to handle unmodeled agent interactions in vehicle platforms.