Deep Reinforcement Learning for Predictive Longitudinal Control of Automated Vehicles

摘要

This paper presents a predictive controller for longitudinal motion of automated vehicles based on Deep Reinforcement Learning. It uses advance information about future speed reference values and road grade changes. The incorporation of this information leads to a new design parameter with a high influence on learning speed: the selection of proper advance knowledge signals during training. We propose a design method which shows improved learning performance in our experiments. The performance of our controller is explored through simulation of a real world driving scenario in a parking garage. We demonstrate that our Reinforcement Learning agent can learn a behavior close to the optimal solution of a Nonlinear Model Predictive Controller, but at reduced computational costs.