Use this link to cite:
https://hdl.handle.net/2183/49543 Sistema multiagente para el control inteligente de semáforos con simulación de tráfico
Loading...
Identifiers
Publication date
Authors
Fernández del Palacio, León
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] La congestión del tráfico urbano afecta diariamente a la economía y la calidad de vida de las ciudades modernas, y el control de los semáforos constituye uno de los mecanismos más directos para mitigarla. La investigación reciente en control semafórico está dominada por el aprendizaje por refuerzo, que funciona adecuadamente en simulación, pero conlleva el uso de información global y de un servidor central, condiciones que rara vez se dan en un despliegue real, además de tratarse de un modelo opaco. Este trabajo aborda el problema desde la viabilidad y propone un sistema multiagente en el que cada intersección es un agente autónomo que decide con información local y se coordina con sus vecinos mediante el protocolo estándar FIPA Contract Net. Sobre este sistema se construye una familia de controladores basados en reglas, evaluados en SUMO sobre redes de complejidad creciente, desde rejillas simples hasta redes reales calibradas, y comparados con métodos clásicos y un controlador de aprendizaje por refuerzo profundo. Los resultados muestran que el valor de la coordinación depende del régimen de tráfico y de la topología, y que el enfoque simbólico es el más viable para un despliegue real.
[Abstract] Traffic congestion is one of the most persistent problems of modern cities, and signal control is one of the most direct mechanisms to mitigate it. Recent research in this field is dominated by reinforcement learning, which performs well in simulation but requires global information and a central server, conditions that are rarely met in a real deployment, in addition to being an opaque model. This work approaches the problem from the standpoint of deployability and proposes a multi-agent system in which each intersection is an autonomous agent that decides with local information and coordinates with its neighbours through the standard FIPA Contract Net protocol. On top of it, a family of rule-based controllers is built and evaluated in the SUMO simulator over a grid network and two real urban networks calibrated with traffic data, against classical baselines and a deep reinforcement learning controller. The results show that the value of coordination depends on the traffic regime and the topology, and that the symbolic approach is the most viable for a real deployment.
[Abstract] Traffic congestion is one of the most persistent problems of modern cities, and signal control is one of the most direct mechanisms to mitigate it. Recent research in this field is dominated by reinforcement learning, which performs well in simulation but requires global information and a central server, conditions that are rarely met in a real deployment, in addition to being an opaque model. This work approaches the problem from the standpoint of deployability and proposes a multi-agent system in which each intersection is an autonomous agent that decides with local information and coordinates with its neighbours through the standard FIPA Contract Net protocol. On top of it, a family of rule-based controllers is built and evaluated in the SUMO simulator over a grid network and two real urban networks calibrated with traffic data, against classical baselines and a deep reinforcement learning controller. The results show that the value of coordination depends on the traffic regime and the topology, and that the symbolic approach is the most viable for a real deployment.
Description
Editor version
Rights
Attribution 4.0 International







