اميرحسين شوروزي

This study addresses the train rescheduling problem in single-track railway networks under disruption conditions, aiming to develop an intelligent real-time traffic management framework. A Reinforcement Learning (RL) approach, specifically the Q-Learning algorithm, is employed as an alternative to classical optimization methods such as Mixed-Integer Linear Programming (MILP), which, despite providing optimal solutions, are computationally expensive an‎d unsuitable for real-time applications. The problem is formulated as a Markov Decision Process (MDP), including the definition of state space, action space, an‎d reward function based on variations in train delays. The learning agent interacts with the environment to minimize the total weighted delay by learning a policy that determines which train should proceed at each conflict point while preventing route conflicts. To eva‎luate model performance, four experimental railway networks of increasing scale (from 5 trains–5 stations to 25 trains–15 stations) were simulated an‎d compared with the MILP benchmark. Results show that in small networks, Q-Learning achieved only 1.3% deviation from the optimal MILP solution, an‎d this gap stabilized at about 5.4% for larger networks. However, the decision-making time of RL remained within milliseconds, up to 37,000 times faster than MILP, whose solution time exceeded ten minutes in large-scale scenarios. This demonstrates that the RL-based model can provide near-optimal, real-time responses, ensuring continuous railway operation even under significant disturbances. Overall, the proposed framework proves that Q-Learning-based rescheduling effectively balances accuracy an‎d computational efficiency, offering a practical an‎d scalable solution for real-time train traffic management. The findings suggest that extending this framework with Deep Reinforcement Learning (DRL) could further narrow the optimality gap while maintaining the exceptional computational speed, enabling advanced decision support systems for modern railway networks.

يادگيري تقويتي , برنامه ريزي مجدد حركت قطار , الگوريتم Q-Learning , زمان‌بندي Real-time , برنامه‌ريزي عدد صحيح مختلط (MILP) , بهينه‌سازي هوشمند

Reinforcement Learning , Train Rescheduling , Q-Learning Algorithm , Real-Time Scheduling , Mixed-Integer Linear Programming (MILP) , Intelligent Optimization

Amirhosein Shorvazi

Masoud Yaghini