Zhao, Zicong (BeiJing JiaoTong University), Xun, Jing (Beijing Jiaotong University), Cao, Yuan (Beijing Jiaotong University), Liu, Jin (University of Leeds)

Towards Explainable Reinforcement Learning for Train Operation: Two Typical Methods

Scheduled for presentation during the Regular Session "S33a-Intelligent Control for Next-Generation Railway Systems" (FR-LM-T33), Friday, November 21, 2025, 11:50−12:10, Southport 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Artificial intelligence (AI) is changing the future of railway, and it is highly expected to improve operation efficiency. As a typical “black-box” system, AI is faced with many risks such as unexplainable results when it is applied to safety-critical systems. This paper proposes two explainable deep reinforcement learning (DRL) methods for train operation optimization. By treating the problem as a multi-classification task, we integrate SHapley Additive exPlanations (SHAP) with deep neural networks (DNN) and GNNExplainer with graph neural networks (GNN) within the soft actor-critic (SAC) framework. These methods explain the important features influencing individual actions and visualize the feature importance and their changes. The proposed approaches address the explainability challenge of DRL in safety-critical systems, enabling transparency in decision-making for train operations.