Deep Reinforcement Learning Based Train Driving Optimization

In contemporary society, railway networks are usually considered among the most critical undertakers of transportation tasks. Thanks to the born-with preferential conditions of the railway systems, in particular the settings and circumstances in which they run, railroad trains are best positioned to become automatic, vis-vis other transportation systems, such as road traffic. Having said that, one must admit that the transformation towards intelligent man-less train driving system will not be easy, given the requirements in high dimensionality, non-linearity, multi-origin restraints, and other factors that may change over time. Challenges, including energy efficiency, on-time performance, and safety, need to be addressed. The authors of the present paper used a deep reinforcement learning (DRL) method to generate energy-optimized train driving solutions. The neural network representation is based on a Bidirectional Long-Short-Term Memory (BLSTM) and an automatic encoder, which integrate sequential properties of the time-series and inherent non-sequential attributes. It is believed that the learned model has value in assisting decision makers in some representative cases of industrial control, as experiments show that this is perfectly adaptable to meet the basic safety, energy saving, and punctuality requirements in train driving.