Smart Transportation Without Neurons -- Fair Metro Network Expansion with Tabular Reinforcement Learning

Researchers demonstrate that tabular reinforcement learning outperforms computationally expensive deep RL methods for metro network expansion problems, achieving 18x fewer training episodes and 12x lower carbon emissions while incorporating fairness criteria. The approach offers an interpretable, resource-efficient alternative to traditional optimization methods for urban transportation planning.

This research addresses a fundamental inefficiency in applying machine learning to combinatorial optimization problems. The Metro Network Expansion Problem represents a class of decision-making challenges where previous approaches either relied on computationally intensive deep reinforcement learning or traditional methods requiring manual constraint engineering. The authors' contribution lies in recognizing that not all complex problems require complex solutions—tabular RL proves sufficient for this domain while dramatically reducing computational overhead.

The broader context reflects growing concern about the environmental and economic costs of training large AI models. As machine learning increasingly influences infrastructure decisions, the sustainability of these systems matters as much as their accuracy. This work aligns with a trend toward efficient AI that questions whether architectural complexity always justifies performance gains. The incorporation of equity metrics into reward functions demonstrates that algorithmic fairness can integrate into optimization without sacrificing efficiency.

For urban planners and transportation authorities, this offers immediate practical value. Cities like Xi'an and Amsterdam served as real-world validation environments, suggesting the method's applicability to existing infrastructure challenges. The 12x reduction in carbon emissions during training translates to measurable environmental benefits, while computational efficiency means smaller cities or agencies with limited resources can implement similar optimization approaches.

The modular framework invites adaptation to related combinatorial problems beyond transportation—logistics networks, utility grids, and telecommunications infrastructure could benefit from similar reformulations. Investors and developers tracking AI infrastructure efficiency should monitor whether this evidence catalyzes broader adoption of lightweight RL methods over deep learning approaches in specialized domains.

• →Tabular reinforcement learning achieves competitive results with 18x fewer training episodes than deep RL on metro expansion problems.
• →The approach reduces carbon emissions by 12x through computational efficiency gains during model training.
• →Reformulating metro expansion as a Non-Markovian Rewards Decision Process enables simpler, more interpretable optimization.
• →Social equity criteria can be directly integrated into reward functions without sacrificing algorithmic performance.
• →The modular, resource-efficient framework offers replicable solutions for other combinatorial optimization problems in infrastructure and logistics.