Talk
Urban traffic disruptions have become a widespread problem in Yerevan over the years.
Overloaded roads with the increasing number of private vehicles have become challenging to
regulate and manage, especially during rush hours. Conventional traffic light systems
typically
operate on fixed-length signal cycles that do not adapt to real-time traffic conditions. This talk
will explore how reinforcement learning can be used to design traffic signals that adapt to
real-time conditions.
The project is based on a simulation of major intersections in Yerevan, including the Komitas corridor, modeled in SUMO (Simulator of Urban Mobility) using publicly available road network data. In this setup, traffic signal control is treated as a reinforcement learning problem, where an agent adjusts signal durations based on observed traffic conditions such as queue length and waiting time.
Rather than focusing on implementation details, the talk will highlight the key modeling decisions
that influence system behavior. This includes how traffic conditions are represented, how control
actions are defined, and how reward functions guide the learning process.
The presentation will show how these choices affect performance in practice. The comparison between
fixed-time control and learned policies under the same traffic conditions will also be discussed to
emphasize where adaptive control improves flow and where it may fall short.
Finally, the talk will highlight the challenges associated with extending this approach from a single intersection to multiple connected ones, where coordination becomes more complex and state definitions become less straightforward. Additional actions required for transitioning from simulation to real-world application will be discussed.
The goal is to provide a practical understanding of how reinforcement learning can be applied to real control problems, and what needs to be considered before moving toward real-world use.
About the Speaker
I am a senior Data Science student at the American University of Armenia, with a particular interest in machine learning and reinforcement learning. I previously spent a semester at Worcester Polytechnic Institute (USA), where I worked on projects in big data analysis and data management.
I am currently working as a Data Scientist at Noble Scripts, where I have worked on data analysis and modeling tasks such as clustering, segmentation, and anomaly detection. More recently, my work has involved multimodal (vision-language) models for video analysis.
My interests are in applying machine learning and reinforcement learning to real-world problems. Recently, I’ve been working on traffic signal control problems using simulation-based approaches.