Adaptive traffic prediction model using Graph Neural Networks optimized by reinforcement learning

Abstract

Traffic prediction is critical for urban planning and transportation management, with significant implications for congestion reduction, resource allocation, and sustainability. Traditional statistical models struggle to capture the complex spatiotemporal dependencies in traffic data, leading to reduced accuracy. Graph Neural Networks (GNNs) have emerged as a more practical approach due to their ability to model these intricate relationships. However, GNN-based methods face challenges in hyperparameter selection, which impacts their performance across diverse traffic scenarios. To address this, this paper proposes an adaptive traffic prediction model that uses reinforcement learning to optimize GNN hyperparameters. Our model significantly reduces the need for manual tuning and improves prediction accuracy across real-world traffic datasets, achieving a 9.8% reduction in Mean Absolute Error (MAE) and 3.6% improvement in Root Mean Squared Error (RMSE) compared to state-of-the-art baselines. In conclusion, dynamic hyperparameter adaptation boosts robustness and efficiency in traffic forecasting. This approach also helps us better understand how to fine-tune GNNs, contributing to the broader knowledge of optimizing machine learning models. Our work helps make traffic prediction more automated and improves how to manage urban transportation.