Publication: Clasificación mediante enjambre de prototipos
Loading...
Identifiers
Publication date
2009
Defense date
2009-07-14
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Los Clasificadores por Vecino ms Prximo (K-NN) han recibido un impulso renovado con la aplicación de metaheursticas de búsqueda (evolutivas, etc.) que permiten optimizar su funcionamiento, mediante la optimización de atributos, reducción de prototipos, y evolución de medidas globales o locales de proximidad; ello permite desarrollar clasificadores robustos e insensibles al ruido, competitivos con otros paradigmas de clasificación. Se aporta a este campo un nuevo algoritmo denominado Clasificador mediante Enjambre de Prototipos (CEP o PSC). Se inspira en el algoritmo de Optimización mediante Enjambres de Partículas (PSO), pero introduce un nuevo enfoque que permite abordar problemas de clasificación de forma flexible y escalable. El algoritmo utiliza un enfoque de Michigan para codificar las posiciones de un conjunto de prototipos, que se desplazan por el espacio de los atributos del problema mediante interacciones de tipo atractivo y repulsivo caractersticas de PSO. Se comporta como un optimizador para una función de evaluación multimodal y dinámica, que mide la calidad de cada prototipo. Los resultados experimentales son competitivos con una variedad de algoritmos de referencia. Como conclusión, se proponen campos de aplicación alternativos, y se avanza la posibilidad de generalizar la propuesta como nueva perspectiva dentro de en la Inteligencia de Enjambre _____________________________________________
Nearest Neighbor Classifiers are a subject of renewed interest with the application of search metaheuristics (such as evolutionary techniques), that may improve their performance by means of attribute selection, prototype reduction and global or local proximity measures; these methods provide the ability to develop robust and noise-insensitive classifiers that may be competitive with other classification paradigms. We contribute to this field a new algorithm, called Prototype Swarm Classifier. It is inspired in the Particle Swarm Optimization algorithm, but it develops a new approach that enables the algorithm to deal with classification problems in a flexible and scalable way. The algorithm uses a Michigan approach to encode positions of a set of prototypes, that move in attribute space using the attractive and repulsive forces that are characteristic of PSO. It behaves by optimization of a multimodal dynamic local fitness function, that measures the quality of each prototype. Experimental results are competitive with those of a set of reference algoritms. In the conclusions, we propose other fields of application, and we introduce the possibility of generalization of the proposal as a new perspective in the field of Swarm Intelligence.
Nearest Neighbor Classifiers are a subject of renewed interest with the application of search metaheuristics (such as evolutionary techniques), that may improve their performance by means of attribute selection, prototype reduction and global or local proximity measures; these methods provide the ability to develop robust and noise-insensitive classifiers that may be competitive with other classification paradigms. We contribute to this field a new algorithm, called Prototype Swarm Classifier. It is inspired in the Particle Swarm Optimization algorithm, but it develops a new approach that enables the algorithm to deal with classification problems in a flexible and scalable way. The algorithm uses a Michigan approach to encode positions of a set of prototypes, that move in attribute space using the attractive and repulsive forces that are characteristic of PSO. It behaves by optimization of a multimodal dynamic local fitness function, that measures the quality of each prototype. Experimental results are competitive with those of a set of reference algoritms. In the conclusions, we propose other fields of application, and we introduce the possibility of generalization of the proposal as a new perspective in the field of Swarm Intelligence.
Description
Keywords
Ingeniería del conocimiento, Inteligencia artificial, Enjambre de prototipos