Edge-Intelligent AIoT-Based Traffic Signal Control Using Deep Reinforcement Learning Under V2X Communication Constraints: A Reproducible Lightweight Surrogate Study

Abstract

Adaptive traffic signal control must account for imperfect sensing and communication, yet reproducible reinforcement-learning (RL) studies often trade physical fidelity for low execution cost. This paper presents an edge-intelligent Artificial Intelligence of Things (AIoT) framework that couples smart-pole observations, a parametric IEEE 802.11p vehicle-to-everything (V2X) channel, and a Double Deep Q-Network (DDQN) controller. The evaluation uses a lightweight research surrogate: a custom Intelligent Driver Model (IDM) car-following implementation, simplified phase and routing logic, and a delay/loss V2X model. It is therefore not a substitute for SUMO/VEINS or field-calibrated validation. Within this stated scope, the artifact evaluates an isolated intersection at three demand levels, a $2{\times }2$ grid, four V2X conditions, five reward settings, and four sensing configurations against five baselines. The archived results show favorable AIoT-DDQN queue management at Medium demand and the highest RL throughput in the four-intersection experiment, whereas Fixed-Time remained stronger in raw throughput under homogeneous Poisson demand. The reported emission quantity is explicitly treated as an idle-plus-acceleration proxy and not as a calibrated estimate of physical pollutants. The study’s contribution is thus a transparent, executable screening framework for controller and communication sensitivity; conclusions about real networks, mixed traffic, or regulated emissions require subsequent SUMO/OMNeT++ and HBEFA/PHEM validation.