TEKNOFEST 2022 — Deep Q-Learning for Traffic Signal Control

Sat, 28 May 2022 00:00:00 +0000

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.

DQN on Rauf Ibishov

TEKNOFEST 2022 — Deep Q-Learning for Traffic Signal Control

What we built #

What we built
#