Privacy-Aware Vehicle Emissions Control System for Traffic Light Intersections

Abstract

This paper proposes a privacy-aware reinforcement learning (RL) framework to reduce carbon emissions of vehicles approaching light traffic intersections. Taking advantage of vehicular communications, traffic lights disseminate their state (i.e., traffic light cycle) among vehicles in their proximity. Then, the RL model is trained using public traffic lights data while preserving private car information locally (i.e., at the vehicle premises). Vehicles act as the agents of the model, and traffic infrastructure serves as the environment where the agent lives. Each time, the RL model decides if the vehicle should accelerate or decelerate (i.e., the model action) based on received traffic light observations. The optimal RL model strategy, dictating vehicles' driving speed, is learned following the proximal policy optimization algorithm. Results show that by moderating vehicles' speed when approximating traffic light intersections, gas emissions are reduced by 25% CO2 and 38% NOx emissions. The same happens for EVs that reduce energy consumption by 20W/h compared to not using the model. at intersections. The final impact of using the model refers to a negligible increment of 20s in the trip duration.