A Decentralized Approach to Autonomous Train Platooning

Benat Froemming-Aldanondo, Maria Gini

Keywords: Multi Agent Reinforcement Learning, Transportation Autonomy, Train Platooning, Decentralized Control Systems

TL;DR: An approach to decentralized train platooning with off-policy DDQN for autonomous distance management and dynamic coupling and uncoupling strategies.

Abstract:

This paper presents a decentralized approach for train platooning, combining reinforcement learning and autonomous decision-making. A Double Deep Q-Network (DDQN) model is employed to ensure safe following within a platoon, while novel algorithms are used to manage real-time decisions for virtual coupling and uncoupling. Our method aims to improve network efficiency by maximizing track usage and
facilitating coordinated, safe platoon formation, leading to increased rail capacity and flexibility. The proposed methods were evaluated in a custom Gym environment, demonstrating scalability and effectiveness in forming platoons.

Submission Number: 6