Publication: Reconocimiento de localizaciones mediante máquinas de soporte vectorial
Loading...
Identifiers
Publication date
2012
Defense date
2012-05-17
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
En la actualidad la inteligencia artificial está presente en prácticamente todos los campos de la vida y aunque es u concepto relacionado con el software es relacionada por todo el mundo con la robótica. El desarrollo de los microprocesadores y de la informática ha posibilitado que las máquinas tengan cierta conciencia de sí mismas y puedan aprender de los actos que realizan para mejorar los procesos para los que fueron construidas. Uno de las especialidades de la inteligencia artificial es el reconocimiento automático, herramientas que permiten gracias a sensores adquirir datos del entorno y reconocerlo. Las formas de abordar el reconocimiento automático son múltiples, una de ellas es mostrarle ejemplos a nuestro sistema (sea un robot, sistema informático…) indicando cuáles de ellos son los deseables y cuales los que no lo son. A base de entrenarlo con ejemplos el sistema se hará una idea o modelo y en el futuro podrá enfrentarse a una situación real y compararla con lo aprendido para saber cómo reaccionar. Esta rutina se conoce como aprendizaje supervisado y es el objetivo de este proyecto. Realizar un sistema informático que simule a un robot que se pueda desplazar por un pasillo reconociendo su emplazamiento. Después de una etapa de aprendizaje se comprobará el grado de fiabilidad en el reconocimiento de las localizaciones para validar el algoritmo. Dentro del concepto de aprendizaje supervisado existen diferentes algoritmos que parten de planteamientos lógicos y matemáticos distintos, siendo quizás el más empleado el de las redes neuronales que simulan la sinapsis del cerebro. En este caso se empleará el de las máquinas de soporte vectorial (SVM, del inglés support vector machine) tipo clasificación. El desarrollo teórico de los distintos métodos de reconocimiento automático y aprendizaje supervisado se explica en el capítulo 2 donde se analizan las diferencias, similitudes y la formulación matemática, para entender los parámetros que afectan a cada algoritmo. Una vez se ha optado por las máquinas de soporte vectorial como base del proyecto, se decide emplear Matlab como soporte informático y la librería Libsvm como algoritmo de aprendizaje y clasificación de SVM. Paralelamente al desarrollo de una interfaz gráfica y al código informático se probó las diferentes opciones de parametrización que tienen las SVM, que van desde la elección de un kernel para la operación interna de datos (y obtención de los vectores soporte) a valores de márgenes y desviaciones. Con el soporte informático y los algoritmos correctamente parametrizados se ensaya con y sin ruido y se valora las debilidades y las fortalezas de la formulación. La parte más importante y objetivo final de un proyecto teórico son las conclusiones ya que aunque los resultados empíricos no sean los deseados, siempre se abren nuevas vías de investigación. En este caso los resultados de las simulaciones fueron positivos y de ellos se extraen conclusiones en el capítulo 5, así como puntos a mejorar de la actual programación y nuevos enfoques que podrían mejorar la robustez y fiabilidad. Finalizando el proyecto aparecen los anexos. Es el espacio reservado a mostrar tabulados los resultados numéricos y las guías de uso de las funciones empleadas en Matlab, tanto las librerías de algoritmos como el código desarrollado en exclusiva para el presente trabajo. _______________________________________________________________________________________________________________________________
Nowadays, the artificial intelligence is present in virtually all fields of life and although it is a software related concept, it is associated worldwide with robotics. The development of microprocessors and computer technology has enabled that machines have some awareness of themselves and can learn from the acts done to improve the processes for which they were built. One of the specialties of the artificial intelligence is the automatic recognition, tool that allows through sensors to acquire data in the environment and recognize it. The approaches to the automatic recognition are many, one of which is to show examples to our system (a robot, computer system ...) indicating which of them are desirable and those which are not. By training it with examples, the system will have an idea or model and in the future it will face a real situation and compare it with what it has learned to know how to react. This routine is known as supervised learning and it is the target of this project. Perform a computer system that simulates a robot that can move down a hallway recognizing its location. After a learning period it shall be tested for reliability in the recognition of the locations to validate the algorithm. Within the concept of supervised learning there are differents algorithms that start from different logical and mathematical approaches, perhaps the most used the neural networks that mimic the brain's synapses. In this case it shall be used support vector machines (SVM) classification type. The theoretical development of the different methods of automatic recognition and supervised learning is explained in Chapter 2 where we analyze the differences, similarities and the mathematical formulation, to understand the parameters that affect each algorithm. Once you have chosen support vector machines as base of the project, we decided to use Matlab as a carrier medium and the LIBSVM library as learning algorithm and SVM classification. Parallel to the development of a graphical interface and the computer code, we tested the different options with the SVM parameter, ranging from the choice of a kernel for the internal operation of data (and obtain the support vectors) to values of margins and deviations. With computer support and properly parameterized algorithms is tested with and without noise and assessed the strengths and weaknesses of the formulation. The most important and ultimate goal of a theoretical project are the conclusions even the empirical results are not as desired, it always open new means of research. In this case the simulation results were positive and they draw conclusions in Chapter 5 as well as areas for improvement of existing programming and new approaches that could improve the robustness and reliability. Completing the project we have the annexes. It is the space to display the numerical results tabulated and guidelines for the use of Matlab functions, both libraries and code algorithms developed exclusively for this proyect.
Nowadays, the artificial intelligence is present in virtually all fields of life and although it is a software related concept, it is associated worldwide with robotics. The development of microprocessors and computer technology has enabled that machines have some awareness of themselves and can learn from the acts done to improve the processes for which they were built. One of the specialties of the artificial intelligence is the automatic recognition, tool that allows through sensors to acquire data in the environment and recognize it. The approaches to the automatic recognition are many, one of which is to show examples to our system (a robot, computer system ...) indicating which of them are desirable and those which are not. By training it with examples, the system will have an idea or model and in the future it will face a real situation and compare it with what it has learned to know how to react. This routine is known as supervised learning and it is the target of this project. Perform a computer system that simulates a robot that can move down a hallway recognizing its location. After a learning period it shall be tested for reliability in the recognition of the locations to validate the algorithm. Within the concept of supervised learning there are differents algorithms that start from different logical and mathematical approaches, perhaps the most used the neural networks that mimic the brain's synapses. In this case it shall be used support vector machines (SVM) classification type. The theoretical development of the different methods of automatic recognition and supervised learning is explained in Chapter 2 where we analyze the differences, similarities and the mathematical formulation, to understand the parameters that affect each algorithm. Once you have chosen support vector machines as base of the project, we decided to use Matlab as a carrier medium and the LIBSVM library as learning algorithm and SVM classification. Parallel to the development of a graphical interface and the computer code, we tested the different options with the SVM parameter, ranging from the choice of a kernel for the internal operation of data (and obtain the support vectors) to values of margins and deviations. With computer support and properly parameterized algorithms is tested with and without noise and assessed the strengths and weaknesses of the formulation. The most important and ultimate goal of a theoretical project are the conclusions even the empirical results are not as desired, it always open new means of research. In this case the simulation results were positive and they draw conclusions in Chapter 5 as well as areas for improvement of existing programming and new approaches that could improve the robustness and reliability. Completing the project we have the annexes. It is the space to display the numerical results tabulated and guidelines for the use of Matlab functions, both libraries and code algorithms developed exclusively for this proyect.
Description
Keywords
Inteligencia artificial, Aprendizaje supervisado, Reconocimiento de formas, Algoritmos computacionales, Simulación por ordenador