A Deep Reinforcement Learning Method for Freight Train Driving Based on Domain Knowledge and Mass Estimation Network

In train marshalling, the mass of the freight train will dynamically change in a wide range, which is the main difficulty in realizing its automatic driving. This paper proposes a deep reinforcement learning method that combines domain knowledge and mass estimation network (MEN). The domain knowledge of excellent drivers is utilized to accelerate the convergence speed of the algorithm and improve the driving performance. Furthermore, the MEN is introduced for estimating the mass of the entire train during driving. Finally, the deep reinforcement learning algorithm selects the output gear based on the estimated mass. The simulation results show that the proposed method has significant effects on performance optimization such as reducing parking error, improving marshalling efficiency, optimizing coupler force and reducing jerk.