Toward Optimal Train Control: An Edge Computing Approach With Adaptive Computation Offloading

Abstract

Train autonomous circumambulate system (TACS) epitomizes a forefront advancement in train control technology, enabling autonomous route triggering, autonomous operation adjustment, autonomous train protection, and autonomous resource management. An essential challenge pertains to the real-time communication and processing capabilities required by the train control systems in TACS. In this article, we propose using edge computing (EC) in TACS to provide real-time communication and computation service for train control. To adapt to the complexity of rail transit operating environment and maintain punctuality and passenger comfort during train operation, we utilize meta-learning to update the traditional train dynamics model and harness the iterative random shooting (IRS) algorithm to optimize the autonomous train control process. Recognizing the limitations of onboard computing capabilities, we propose a model-based meta reinforcement learning approach to obtain the optimal task offloading policy. The optimal policy evaluates the channel conditions and computing resource utilization of onboard devices, and wisely determines whether to perform local computing or transmit data to EC devices for processing. In addition, our approach uses meta reinforcement learning to train the environment dynamics model prior such that, when combined with recent data, this prior can be rapidly adapted to the local environment. The model-based meta reinforcement learning approach is quite suited for the urban rail transit system where different rail lines have different operating environments, and we do not have enough data to finish a regular training task. Empirical evidence demonstrates that our proposed framework furnishes the train autonomous control system with reliable and real-time computing services, thereby significantly enhancing operational efficiency through our novel adaptive computation offloading policy.