A Deep Reinforcement Learning-Based Approach to Intelligent Powertrain Control for Automated Vehicles

Abstract

The development of a powertrain controller for automated driving systems (ADS) using a learning-based approach is presented in this paper. The goal is to create an intelligent agent to learn from driving experience and integrate the operation of each powertrain unit to benefit vehicle driving performance and efficiency. The feature of reinforcement learning (RL) that the agent interacts with environment to improve its policy mimics the learning process of a human driver. Therefore, we have adopted the RL methodology to train the agent for the intelligent powertrain control problem in this study. In this paper, a vehicle powertrain control strategist named Intelligent Powertrain Agent (IPA) based on deep Q-learning network (DQN) is proposed. In the application for an ADS, the IPA receives trajectory maneuver demands from a decision-making module of an ADS, observes the vehicle states and driving environment as the inputs, and outputs the control commands to the powertrain system. As a case study, the IPA is applied to a parallel hybrid vehicle to demonstrate its capability. Through the training process, the IPA is able to learn how to operate powertrain units in an integrated way to deal with varied driving conditions, and vehicle’s speed chasing ability and power management are improved along with its driving experience.