Transformer-Based Reinforcement Learning for Adaptive Traffic Signal Control in Smart City Environments to Reduce Congestion and Emissions

Rapid urbanization and escalating vehicular volumes present critical challenges to modern metropolitan mobility, exacerbating street-level congestion and driving up transportation-related greenhouse gas emissions. Traditional adaptive traffic signal control heuristics frequently fail to capture high-order, non-linear spatio-temporal correlations across interconnected arterial networks. In this study, we propose a novel Spatio-Temporal Transformer-based Multi-Agent Reinforcement Learning (ST-TARL) framework tailored for decentralized adaptive traffic signal control in smart city ecosystems. By integrating self-attention mechanisms across both temporal sequence horizons and dynamic inter-intersection spatial graphs, each signal agent learns cooperative control policies that explicitly optimize vehicular throughput while mitigating stop-and-go driving patterns. Furthermore, we incorporate an emissions-aware multi-objective reward formulation derived from instantaneous vehicular energy consumption dynamics. Extensive microscopic traffic simulations conducted in SUMO (Simulation of Urban MObility) across both synthetic grid topologies and real-world urban road networks demonstrate that ST-TARL significantly outperforms state-of-the-art baselines. Specifically, our model achieves a 26.4% reduction in average vehicular delay, a 31.8% decrease in cumulative intersection queue lengths, and an 18.2% drop in tailpipe carbon dioxide (CO2) emissions compared to leading deep multi-agent reinforcement learning approaches. These findings underscore the viability of transformer-driven policy optimization as an enabling methodology for sustainable, responsive, and eco-friendly urban intelligent transportation infrastructures.