Multi-modal and multi-agent reinforcement learning framework for urban traffic flow prediction and signal control optimization

Abstract

The rapid urbanization of cities has exacerbated traffic congestion, resulting in significant environmental impacts, including elevated greenhouse gas emissions and deteriorating air quality. Traffic management systems, while effective in many contexts, often fail to consider the ecological and dynamic complexities of modern urban environments. This paper introduces MM-STMAP, a framework for urban traffic management that integrates multi modal perception with deep reinforcement learning. The approach utilizes a spatio temporal graph convolutional network to model intricate traffic patterns across diverse urban environments, while incorporating real-time environmental data, including meteorological factors, to address the ecological limitations of traditional traffic systems. A linear attention mechanism is employed to optimize computational efficiency in processing large-scale, dynamic traffic data, thereby enhancing both operational performance and energy consumption. The multi agent reinforcement learning structure governs the coordination of traffic signals across intersections, achieving a dual optimization of reduced vehicular delays and minimized emissions. Empirical evaluations on major metropolitan datasets demonstrate that MM-STMAP outperforms existing traffic management methods and significantly enhances traffic flow efficiency. The model's ability to integrate heterogeneous data streams spanning traffic sensors and environmental reports enables a comprehensive and adaptive approach to urban mobility, supporting the development of sustainable smart city infrastructure.