Adaptive Traffic Signal Control at a Single Intersection Using Deep Reinforcement Learning: A SUMO-Based Comparative Study

Urban road networks lose a large amount of time and fuel to congestion at signalised intersections, and conventional fixed-time controllers cannot react to the changing and uneven traffic demand seen in practice. Reinforcement learning (RL), and in particular deep reinforcement learning (DRL), has been proposed as a way to learn signal-timing policies directly from interaction with traffic. This paper proposes a simulation-based framework for adaptive traffic signal control at a single four-way intersection, built on the SUMO microscopic traffic simulator. Two DRL agents, a Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), are formulated as Markov decision processes with a queue- and waiting-time-based state, a phase-selection action space, and a reward based on the change in cumulative vehicle waiting time. The framework defines a comparative evaluation protocol in which both agents are measured against fixed-time, vehicle-actuated and max-pressure controllers across five demand scenarios, using average delay, queue length, stops, and throughput over multiple random seeds. This is a proposal paper: the framework is specified in full, but no experiments have been run, so no results are reported. The paper also states the expected outcomes, limitations, and directions for implementation.