The Restricted Newton Method for Fast Nonlinear Model Predictive Control

We introduce the Restricted Newton’s Method (RNM), a basic optimization method, to accelerate model predictive control turnaround times. RNM is a hybrid of Newton’s method (NM) and gradient descent (GD) that can be used as a building block in nonlinear programming. The two parameters of RNM are the subspace on which we restrict the Newton steps and the maximal size of the GD step. We present a convergence analysis of RNM and demonstrate how these parameters can be selected for MPC applications using simple machine learning methods. This leads to two parameter selection strategies with different convergence behaviour. Lastly, we demonstrate the utility of RNM on a sample autonomous vehicle problem with promising results.