HYBRID SENSING AND CLOUD DEPLOYED AGENT MESH FOR AI‑DRIVEN TRAFFIC OPTIMIZATION IN CONGESTED CITIES

Urban traffic congestion in crowded cities is increasingly shaped by short term demand fluctuations, spillback formation, and heterogeneous vehicle dynamics, which challenge fixed time signal plans and reactive controllers. This paper proposes an AI enabled intersection agent that combines (i) a deep learning (DL) predictor for short horizon traffic forecasting (15–60 minutes) and (ii) an intersection level machine learning (ML) controller that applies bounded adaptations of green times based on predicted demand and real-time sensing. To support both model training and closed loop evaluation without requiring immediate access to city scale labeled data, we introduce a two simulator workflow: a correlation aware city behavior generator that synthesizes realistic temporal patterns (e.g., day/night and weekday cycles, holidays, weather dependent modal shifts, and long term trends) to produce training data for the DL predictor, and a microscopic grid simulator that models heterogeneous vehicles (AUTO, VAN, BUS) and intersection geometry to quantify control impact under congestion. The evaluation uses a simplified Intersection Type‑1 sensing layout for results reporting and discusses extensibility to Intersection Type‑2 designs, including full control and simplified sensor configurations suitable for different deployment constraints. Performance is summarized through an Influence Coefficient (CI) that measures the control induced change in congestion factor between baseline and ML enabled runs. Simulator experiments indicate that the proposed agent can improve congestion outcomes while remaining implementable under cost constrained sensing, and we outline cloud/MLOps requirements for periodic retraining, governance, and safe deployment in real world urban infrastructures. This correlation aware synthetic data generator extends the dynamic historical data generation concept previously introduced for traffic management model training (Author, 2022).