A Hybrid Machine Learning-based Control Strategy for Autonomous Driving Optimization

: Developing autonomous vehicles is a highly important topic in the field of intelligent transportation systems. Automated steering is a crucial function in the autonomous vehicle. Therefore, it is urgent to either develop a new effective control strategy or improve existing ones. A variety of control strategies are used for this purpose, most with limitations related to their computing capabilities with the highly complex systems or to lack of efficacy related to maintaining the balance between driving performance and driving smoothness. In this paper, three different machine learning-based models were developed to perform an autonomous driving task: a supervised learning model (Deep Neural Network, DNN), a reinforcement Deep Q-learning model (DQN), and a hybrid model. The DNN model was trained based on the behavior of the classical MPC controller. The DQN was designed with the same structure as the DNN and trained by directly interacting with the driving environment. The hybrid model is a combination of supervised and reinforcement learning algorithms, where the trained DNN model is used as a decision-maker (Actor) in a deep deterministic policy gradient reinforcement learning model. The behavior of the designed models was compared based on several performance indicators, including the ability to drive the vehicle along the desired trajectory, the response time, and the smoothness of the driving system. The results show that the DNN model was able to imitate the behavior of the traditional MP Controller efficiently and all three machine learning models successfully drive the vehicle along the desired path. The hybrid model achieves the best results and improved the smoothness of the driving system with a reasonable response time.