Author(s) :

Shaunak Thamke.

Volume/Issue :

Abstract :

Traffic congestion remains a major challenge in modern cities. It contributes to delays, fuel waste, pollution, and reduced quality of life. Most intersections still rely on fixed-time signals that do not adapt to real-time traffic demand. This research investigates whether Deep Reinforcement Learning (DRL) can optimize signal timing more effectively than traditional approaches. Historical Kaggle traffic data and real-time New York City (NYC) data was used to generate Poisson-based vehicle arrivals across fourway intersections. Two DRL agents, Proximal Policy Optimization (PPO) and Deep Q Network (DQN), were trained to minimize vehicle wait-time (delays) and queue length (congestion). Although DQN has been widely used in previous traffic-signal studies, PPO remains underexplored. This work provides an evaluation of both methods against two baselines: Fixed-Time Baseline Controller (FBC) that used Kaggle data and Analytical Baseline Controller (ABC) that used real-time NYC data. PPO performed the best and reduced average wait-time by 90.9% compared to FBC and 69.46% compared to ABC. DQN saved 61.32% compared to ABC. PPO consistently delivered greater reductions and stability across runs. Generated heatmaps based on 20 simulations confirmed the adaptability of PPO in maintaining low queue lengths and avoiding severe congestion, common in fixed-time and analytical systems. The DQN heatmap showed reduced queues with occasional variability, while PPO was stable. These findings highlighted the potential of PPO for dynamic, data-driven traffic control which was demonstrated using interactive traffic simulation. This research has great potential in improving traffic flow efficiency, reducing congestion and pollution.