Meta-Reinforcement Learning for Sustainable and Self-Optimizing Smart Cities

Abstract

Smart cities are increasingly challenged by dynamic urban environments, heterogeneous data sources, and sustainability constraints that demand intelligent, adaptive, and autonomous decision-making systems. Conventional reinforcement learning (RL) approaches often suffer from slow convergence and limited generalization when exposed to rapidly changing urban scenarios. To address these limitations, this paper presents a Meta-Reinforcement Learning (Meta-RL) framework for sustainable and self-optimizing smart cities, enabling rapid adaptation across diverse urban tasks such as traffic control, energy management, waste optimization, and public safety monitoring. By learning a meta-policy that captures transferable knowledge across multiple city environments, the proposed framework significantly improves learning efficiency, scalability, and robustness under non-stationary conditions. Sustainability objectives—including energy efficiency, carbon footprint reduction, and quality-of-service (QoS) optimization—are explicitly embedded into the reward structure to ensure environmentally responsible decision-making. Experimental evaluations across simulated smart-city scenarios demonstrate that the Meta-RL approach outperforms traditional RL and heuristic methods in terms of convergence speed, resource utilization, and long-term sustainability metrics. The results highlight the potential of Meta-RL as a foundational intelligence layer for next-generation autonomous and sustainable urban ecosystems.