Graph neural networks and reinforcement learning in traffic optimization in cities

09:55-10:10, January 28

Talk/ Overview

We will present recent results of the TensorCell research project aiming to optimize complex processes such as vehicular traffic in large cities.First, we will show our current approach in which we train graph neural networks and LightGBM to approximate outcomes of microscopic traffic simulations (performed using Traffic Simulation Framework runs on the whole road network of Warsaw) in order to evaluate efficiently the quality of traffic signal settings. Thanks to that, we are able to use machine learning models as surrogate models of time-consuming simulations and apply metaheuristics (e.g., genetic algorithms, simulated annealing) to find heuristically optimal settings of traffic signals.Second, we will present reinforcement learning approaches. We implemented a simple traffic simulator based on a cellular automaton model (Nagel-Schreckenberg model) and we are testing whether traffic signal controllers may learn how to manage traffic optimally on a single road corridor with few intersections.Finally, we will briefly outline future directions of the research and invite potential collaborators to the project.

Talk/ Speakers

Kamil Kaczmarek

Product Owner, neptune.ml & Reinforcement Learning ResearcherTensorCell

AMLD / Global partners