Practice of reinforcement learning-based traffic flow optimization strategies for intelligent traffic management system applications
Abstract
As urbanization and industrialization escalate, the burgeoning traffic flow precipitates grave issues of congestion, environmental degradation, and heightened energy consumption, thereby imperiling the sustainable evolution of urban centers and the life quality of their denizens. The inception of Intelligent Transportation Systems (ITS) is designed to enhance the operational efficiency and efficacy of traffic management via the deployment of sophisticated information technology, data analytics, and machine learning methodologies. In this treatise, we introduce a pioneering deep learning architecture, denoted as RL-LGCN, for the short-term forecasting of traffic flow, integrating Graph Convolutional Networks (GCN), Gated Recurrent Units (GRUs), and Reinforcement Learning (RL). Initially, we undertake the preprocessing of traffic data, encompassing data purification and the imputation of missing values, followed by the extraction of periodic attributes using the trigonometric encoding technique to ready the data for model input. Subsequently, the spatial interdependencies within the traffic network are delineated by GCN, while the extensive temporal dependencies of the time-series data are managed by GRU. Building on this foundation, the model undergoes continuous training and refinement by perpetually updating the interplay between the model and the external milieu through reinforcement learning. Our empirical evaluations reveal that the predictive accuracy of the RL-LGCN framework is markedly superior on both public and proprietary datasets, outstripping the performance of conventional monolithic models. This research furnishes robust technical backing for intelligent traffic management systems, propels the advancement of traffic flow prediction techniques, and establishes a cornerstone for the future's intelligent and automated urban transportation infrastructures.