PM
Back to blog
M.S. thesis noteReinforcement Learning20217 min read

An Intelligent Traffic Light Control System Using Deep Neural Networks

My 2021 M.S. thesis used Reinforcement Learning with Deep Q-Learning, SUMO simulation and YOLO-based traffic estimation for adaptive traffic-signal control.

Authors: Pouria Maleki
Pouria MalekiBu-Ali Sina University · M.S. ThesisBlog updated: 2026-08-14
An Intelligent Traffic Light Control System Using Deep Neural Networks — research article by Pouria Maleki
Key results & takeaways
SUMO provides the traffic environment while Deep Q-Learning drives sequential signal decisions
Computer vision supplies traffic-state estimates for adaptive control
The project connects perception, simulation and control rather than treating AI as an isolated model

Research visuals & results

2 figures
Animated SUMO intersection environment used for traffic-signal-control experiments.
Animated SUMO intersection environment used for traffic-signal-control experiments.
Four-way intersection and traffic-phase representation.
Four-way intersection and traffic-phase representation.

Traditional traffic lights usually operate with fixed timing plans or manually tuned rules. My M.S. thesis explored a more adaptive idea: treat the intersection as a reinforcement-learning environment and let an intelligent agent learn which signal action reduces congestion.

The learning loop

The traffic environment was modeled in SUMO. At every decision step, the agent received a representation of current traffic conditions, selected a signal action and observed a reward related to traffic performance.

text
Traffic state → Deep Q-Learning agent → Signal phase
      ↑                               ↓
      └──────── reward / delay ───────┘

Connecting computer vision

A practical adaptive controller needs a reliable estimate of traffic density. The project therefore investigated real-time video processing and YOLO-based vehicle detection as the perception layer.

This creates an end-to-end architecture:

  1. Camera observes the intersection
  2. Detector estimates vehicles and traffic density
  3. State is sent to the RL agent
  4. Agent chooses the next traffic-light action
  5. The environment produces a new state and reward

Why reinforcement learning fits traffic control

Traffic is stochastic and highly dynamic. The optimal decision depends not only on the current queue but on how current actions influence future congestion. Reinforcement learning is naturally suited to this sequential decision problem.

Transferable insight

The most important lesson from the thesis was learning how to combine mathematical modeling, simulation, perception and sequential optimization in one engineering system. That same pattern appears in robotics, resource allocation and many real-time AI applications.

Related research