Nonlinear Model Predictive Control Reduction Using Truncated Single Shooting

Anson Maitland, Mohit Batra, John McPhee

Abstract: We present a new scheme to reduce the computational burden of Nonlinear Model Predictive Controllers (NMPCs). Working with the direct single shooting formulation of NMPC, we identify the timestep as a `small' parameter and conduct a perturbative analysis of Newton's step. This analysis leads to the introduction of a truncated Lagrangian that can be used to form a new quasi-Newton method with a symbolically reduced Hessian for online optimization. To the authors' knowledge, these methods have not been presented before. In this work we leverage the power of symbolic computation to generate efficient controller code for online evaluation. The reduction methods were applied to diesel engine control and electric vehicle cruise control problems and achieved turnaround times more than 2 times faster with no tradeoff in controller performance. These initial results are quite promising for the development of real-time NMPCs and introduce some new avenues for research.