↓ Skip to main content
Rauf Ibishov
Rauf Ibishov
  1. Blog/

TEKNOFEST 2022 — Deep Q-Learning for Traffic Signal Control

·2 mins
Python Reinforcement Learning Deep Q-Learning DQN Competition Hackathon
Rauf Ibishov
Author
Rauf Ibishov
Based in Heilbronn, Germany. Building retrieval and ranking systems at Azerbaijan’s National AI Center. Starting my MSc at TUM Heilbronn — looking for werkstudent roles in NLP, search, or ML infrastructure.
LeetCode

Key lesson: reward shaping and state representation matter more than the choice of algorithm — a well-designed reward with a simple DQN beats a fancy algorithm with a poor reward function.

In May 2022, our team competed among 40 teams at the Smart Qarabag Hackathon, part of the TEKNOFEST Azerbaijan international technology festival. We built a reinforcement learning model for intelligent traffic signal control.

What we built
#

The core idea: treat a traffic intersection as an environment and the signal controller as an agent. We used Deep Q-Learning to train the agent to optimize signal timing — learning when to switch lights based on real-time queue lengths, waiting times, and traffic density across all directions.

The MDP we set up was deliberately compact:

  • State — traffic density, queue lengths, and waiting times per direction.
  • Action — signal phase selection plus the duration of each phase.
  • Reward — a function of waiting-time reduction and throughput improvement.

The model controlled a single intersection, adjusting green/red phase durations to minimize average vehicle waiting time and maximize throughput. Rather than following fixed-cycle timers (which is what most real intersections still do), the agent learned adaptive policies that responded to actual traffic conditions.

Our longer-term vision was multi-intersection coordination — having neighboring intersection agents communicate to achieve optimal traffic flow across a network, not just locally. We scoped the hackathon demo to a single intersection to keep it demonstrable, with the coordination layer as a clear next step.

What came out of it
#

The project caught the attention of an AI specialist at the event who offered us training and mentorship to develop the system further. It was a good validation that the approach had real-world potential beyond the hackathon setting.

This was also my first hands-on project with reinforcement learning. The gap between “understanding Q-learning from a textbook” and “making an agent actually converge on a useful policy in 48 hours” was humbling — reward shaping and state representation turned out to matter far more than the choice of algorithm.

  • Event: TEKNOFEST Azerbaijan 2022 — Smart Qarabag Hackathon
  • Scale: 40 teams
  • Tech: Deep Q-Learning, Python

May 2022 — Baku, Azerbaijan

Related#