Real World Data-Based Deep Reinforcement Learning for Traffic Management and Emissions Reduction in a Low Emission Zone

Abstract

ABSTRACT : In recent years, vehicular congestion has become one of the major issues for large cities, as it brings about a series of negative consequences that affect both the economy of the cities and the health of their residents. In economic and environmental terms, the high costs associated with traffic congestion result from increased fuel consumption, greater wear and tear on road infrastructure, disruptions to public transportation services, and the deterioration of air quality due to the emission of pollutants. This leads to an increase in respiratory and cardiovascular diseases that impact various population groups.

For example, the cities that make up the Metropolitan Area of the Aburrá Valley (AMVA - Área Metropolitana del Valle de Aburrá), in the state of Antioquia (Colombia), are not immune to the effects that the transportation sector has on air quality (mainly in the capital, Medellin). According to the entity, in the emissions inventory for the base year 2018, 91% of PM2.51 emissions came from mobile sources such as trucks, buses, 4-stroke motorcycles, and private vehicles. Given the relationship between transportation and environmental pollution, major cities like Rome, Milan, and London have started to engage in more conscious mobility planning through strategies such as higher parking fees to encourage the use of public transportation, standardization of speeds for specific urban areas, the establishment of restricted traffic zones during specific hours, and the labeling of Low Emission Zones (LEZ). Moreover, improvements have not only been sought through policies and infrastructure but also through academic and private sector research and development of traffic management systems. From these research and experiments, strategies heavily reliant on technological and computational inputs have emerged, such as route planning, Vehicle-to-Infrastructure (V2I) communication, and traffic light control, taking into account weather conditions, traffic history, and information collected from sensors and communication systems. Currently, with the rise of artificial intelligence (AI), various machine learning techniques have been explored as potential solutions that, in conjunction with the aforementioned strategies, could drive the development of more robust and dynamic systems capable of adapting to traffic conditions. Among these techniques are traffic prediction models, traffic lights optimization, traffic simulation, reinforcement learning, and computer vision. This research work aims to address traffic management from a multi-agent perspective, considering the existing road infrastructure in the city of Medellin, to execute traffic light control in a simulated environment and evaluate the impact that an intelligent management system could have on the city’s traffic, which could positively affect the air quality in the study region.