An MPC Approximation Approach for Adaptive Cruise Control With Reduced Computational Complexity and Low Memory Footprint

This work demonstrates the application of deep neural networks (DNN) to alleviate the computational complexity associated with Model Predictive Control (MPC), which has always been an obstacle hindering the practical adoption of MPC. This challenge is particularly critical in applications for autonomous vehicles where achieving multiple objectives while enforcing a certain number of system constraints is essential. We first revisit and design a control algorithm tailored to the Adaptive Cruise Control (ACC) problem. The developed algorithm consists of two distinct implicit MPCs, each addressing a specific control mode, namely velocity and space control. Multiple control objectives and constraints are integrated into the algorithm synthesis to ensure satisfactory control performance. We further adopt supervised learning with deep neural networks to reduce the computational cost of MPC, thereby making MPC more accessible for practical use. Simulation results affirm that the DNN-based approximated policy can match the control performance in terms of both tracking precision and constraint satisfaction of state-of-the-art solvers dedicated to optimization problems. Remarkably, the execution time of the approximated policy is approximately one order of magnitude lower than that of implicit MPCs, while its memory footprint is significantly lower than those of its counterparts, thereby emphasizing its distinct advantages.