Title: Neuron-Free Smart Transit: Achieving Equitable Metro Expansion via Tabular Reinforcement Learning

Abstract

This study addresses the Metro Network Expansion Problem (MNEP), a specific component of the broader Transport Network Design Problem (TNDP) aimed at extending metro infrastructure to meet passenger demand. Conventional techniques for solving MNEP depend on exact algorithms or heuristics, which necessitate manually defined constraints to narrow the search space. While Deep Reinforcement Learning (Deep RL) has recently gained traction for its proficiency in handling complex, sequential decision-making tasks, it suffers from high computational costs, significant environmental impact, and a lack of interpretability that demands extra engineering effort.

We argue that the scale of MNEP challenges does not justify the use of Deep RL. By reframing the problem as a Non-Markovian Rewards Decision Process (NMRDP), we apply tabular reinforcement learning to attain performance levels comparable to Deep RL, but with a drastic reduction in the number of required training episodes and enhanced clarity. Furthermore, we integrate social equity metrics into the reward structure, prioritizing both efficiency and fairness, which demonstrates the adaptability of our proposed method.

Testing our approach in real-world scenarios in Xi’an and Amsterdam reveals that it cuts the total number of training episodes by 18 times and lowers total carbon emissions by 12 times, on average, while maintaining competitiveness with Deep RL models. This strategy provides a solution that is modular, interpretable, resource-efficient, and easily replicable, with promising implications for other combinatorial optimization challenges.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC