Publication: Control inteligente de semáforos mediante aprendizaje automático
Loading...
Identifiers
Publication date
2016
Defense date
2016-10-10
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Este trabajo se centra en el control de semáforos de forma inteligente. Se presentan varios sistemas funcionales en la actualidad que utilizan diversas técnicas de Inteligencia Artificial para optimizar la gestión de tráfico, desde árboles de decisión y redes de neuronas hasta algoritmos genéticos.
Otra técnica de la Inteligencia Artificial altamente expandida en este campo es el aprendizaje por refuerzo. Gracias a esta técnica, los semáforos son entrenados inmersos en un entorno específico de tráfico y son capaces de aprender a través de la experiencia.
En este proyecto se presenta un sistema multi-agente distribuido, donde se aplica aprendizaje por refuerzo (Q-learning) para resolver el problema de la gestión de tráfico. La particularidad de este sistema reside en que cada agente controla solamente un semáforo, basándose en la información local del mismo, como el número de vehículos o la velocidad media en su carril.
Este trabajo se desarrolla en el simulador SUMO. El sistema interactúa con SUMO a través de la interfaz TraCI, que ofrece todas las herramientas necesarias para controlar la simulación a la vez que pone a disposición del usuario todas las métricas mencionadas anteriormente.
Para evaluar el sistema desarrollado se ejecuta en varios escenarios, un sistema de política fija y un sistema de política aleatoria. Cada escenario dispone de un mapa y un nivel de tráfico distinto, con el objetivo de comprobar el comportamiento de cada sistema con tráfico fluido y congestionado. En base a los resultados que estos experimentos arrojan, se concluye que el sistema desarrollado gestiona de forma más eficiente el tráfico congestionado que el de política fija, aunque es menos eficiente en situaciones de tráfico fluido.
This project is about creating a system that manages traffic efficiently. Big cities have some serious troubles when it comes to traffic jams and this has a big impact on both the economy and the enviroment. The study The future economic and environmental costs of gridlock in 2030 made by the CEBR (Centre for Economics and Business Research) concludes that the total cost of traffic jams in just four big countries (France, Germany, Great Britain and the United States) raises up to 200 billion dolars in 2013. But this can be lower if an intelligent system is created and implemented. The current system that is installed all around the world is a centralized system that contains traffic lights that only changes in a fixed interval. This system has the exact same behavior regardless of the situation of the traffic and it fails at avoiding traffic jams. In the recent past, some researchers have created intelligent traffic lights that make decisions based on the traffic that exists in that exact moment, improving the traffic flow and reducing the jams. But these systems are generally centralized, which means they are not scalable. In this project, we will try to create a distributed multi-agent system using machine learning, where each traffic light will be independent and will only have information about the lane they are controlling. This way, each traffic light will learn to make better decisions and the system will be highly scalable due to the possibility of installing it anywhere, regardless of where it is situated. This system will be integrated in the simulator chosen to do the tests. This simulator is SUMO, a traffic simulator that offers a realistic vehicle behavior and an easy interface called TraCI. This interface allow us to control the simulation, change any traffic light at will and get any information about number of vehicles or their speed in any lane. It also supports many programming languages: Python, Java, Matlab and C++. It has some disadvantages like the inability to rewind time, but it is the best simulator to use in this case. As it was said previously, machine learning will be used in the system that it is going to be created. Machine learning is a subfield in computer science which objective is to develop some techniques to give computers the ability to learn. These techniques are about generating behaviours from examples and they focus on solving classification, optimizations and decision-making problems. Some of its applications are: detecting credit card fraud, speech and handwriting recognition, robot locomotion and medical diagnosis. Reinforcement learning is an area of machine learning that it is based on a software agent that must learn a behavior through trial and error, interacting with an enviroment so as to maximize a reward. The reinforcement model consists of: A set of states representing the enviroment at a given time. A set of actions that the agent can perform. A reward function that defines the reward received by the agent after performing an action.
This project is about creating a system that manages traffic efficiently. Big cities have some serious troubles when it comes to traffic jams and this has a big impact on both the economy and the enviroment. The study The future economic and environmental costs of gridlock in 2030 made by the CEBR (Centre for Economics and Business Research) concludes that the total cost of traffic jams in just four big countries (France, Germany, Great Britain and the United States) raises up to 200 billion dolars in 2013. But this can be lower if an intelligent system is created and implemented. The current system that is installed all around the world is a centralized system that contains traffic lights that only changes in a fixed interval. This system has the exact same behavior regardless of the situation of the traffic and it fails at avoiding traffic jams. In the recent past, some researchers have created intelligent traffic lights that make decisions based on the traffic that exists in that exact moment, improving the traffic flow and reducing the jams. But these systems are generally centralized, which means they are not scalable. In this project, we will try to create a distributed multi-agent system using machine learning, where each traffic light will be independent and will only have information about the lane they are controlling. This way, each traffic light will learn to make better decisions and the system will be highly scalable due to the possibility of installing it anywhere, regardless of where it is situated. This system will be integrated in the simulator chosen to do the tests. This simulator is SUMO, a traffic simulator that offers a realistic vehicle behavior and an easy interface called TraCI. This interface allow us to control the simulation, change any traffic light at will and get any information about number of vehicles or their speed in any lane. It also supports many programming languages: Python, Java, Matlab and C++. It has some disadvantages like the inability to rewind time, but it is the best simulator to use in this case. As it was said previously, machine learning will be used in the system that it is going to be created. Machine learning is a subfield in computer science which objective is to develop some techniques to give computers the ability to learn. These techniques are about generating behaviours from examples and they focus on solving classification, optimizations and decision-making problems. Some of its applications are: detecting credit card fraud, speech and handwriting recognition, robot locomotion and medical diagnosis. Reinforcement learning is an area of machine learning that it is based on a software agent that must learn a behavior through trial and error, interacting with an enviroment so as to maximize a reward. The reinforcement model consists of: A set of states representing the enviroment at a given time. A set of actions that the agent can perform. A reward function that defines the reward received by the agent after performing an action.
Description
Keywords
Inteligenica artificial, Control del tráfico, Algoritmos genéticos, Sistema multi-agente distribuido