Une approche coopérative basée sur l’IoT pour améliorer la qualité du trafic routier
No Thumbnail Available
Date
2023-12-21
Authors
Tarek Amine HADDAD
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Nowadays, automobiles have become very useful for the daily transportation of both people and goods. However, the increase in their numbers has generated a significant rise in demands for the use of road networks, which can lead to delays, traffic congestion, and poor traffic flow, especially in large cities and global metropolises. This type of problem often referred to as road congestion, can be caused by other factors such as road works, accidents, insufficient traffic management, etc. Many strategies to reduce traffic congestion have been adopted by governments and transport agencies.
These strategies include building new roads, planning additional public transport routes, real- time traffic management, etc.
On the other hand, traffic signal control (TSC) represents one of the modern trends that can play a crucial role in traffic management. It also refers to the management and coordination of traffic lights on road networks in order to effectively control traffic and improve traffic flow. This can involve adjusting the traffic light cycle time based on real traffic to optimize certain performance parameters. Modern TSC systems can adopt technologies such as traffic sensors, surveillance cameras and data processing algorithms to optimize traffic management.
Reinforcement learning (RL) is an intelligent approach widely adopted by adaptive TSC systems to optimize traffic signal management, enhancing traffic flow. Indeed, it is possible to train a system to learn how to adjust traffic light cycle times based on the state of real traffic in order to reduce road congestion.
In this thesis, we propose several cooperative approaches based on Deep Reinforcement Learning (DRL) to intelligently optimize the management of traffic lights in a road network with multiple intersections. We have thus modeled our problem as a multi-agent reinforcement learning system (MARL). This involves the use of multiple agents each of whom can learn to make decisions in terms of adjusting light cycle times to the local traffic situation, and these decisions can be synchronized with the decisions made by other agents to ensure optimal functioning of the entire system. In such approaches, each agent can receive from its neighbors their states, actions and rewards,
combining them with its own state, action and reward to make the appropriate decisions.
Experimental results under different scenarios show that the proposed approaches outperform many state-of-the-art approaches in terms of three parameters: Average Waiting Time (AWT), Average Queue Length (AQL) and average CO2 emission (AEC).