Neuro-VAE-Symbolic Dynamic Traffic Management

Abstract

Modern traffic management must balance the ability to predict complex, non-stationary flow patterns with the strict enforcement of safety and legal constraints (e.g., minimum pedestrian intervals, maximum cycle lengths). While generative models such as Variational Autoencoders (VAEs) can effectively capture traffic dynamics, purely data-driven approaches often fail to respect domain-specific rules. We propose a Neuro-VAE-Symbolic framework that combines the flexibility of VAE-based generation with the rule-enforcing capability of symbolic reasoning. By projecting raw neural outputs onto a feasible action space defined by traffic regulations, our method ensures both adaptability to dynamic demand and strict constraint compliance. Experiments on benchmark datasets show that our method reduces average waiting time by up to <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">15%</b>, cuts queue lengths by <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10–20%</b>, and increases throughput by <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">5–8%</b>, all while maintaining near-zero rule violations. These results demonstrate that neuro-symbolic integration enables high-fidelity traffic scenario generation alongside reliable, regulation-abiding decision-making, offering a robust solution for dynamic urban traffic control.