Paralelización de la meta-heurística ¨Optimización en Mallas Variables¨ utilizando la Arquitectura CUDA
Archivos
Fecha
2015-06-24
Autores
Martínez Prieto, Adrian
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
En el campo de la optimización, tanto académico como industrial los problemas que se presentan a diario son a menudo complicados y NP-difícil. La solución computacional de los mismos no puede llevarse a cabo de una manera exacta en un plazo de tiempo razonable, y el uso de recursos de hardware de cómputo para esta solución aumenta considerablemente.
Para lidiar con tal asunto, el diseño de los métodos de solución debe estar basado en el uso conjunto de los enfoques avanzados de la optimización combinatoria, los métodos de paralelismo e ingeniería a gran escala. Por lo que existen diferentes ramas de la investigación dedicadas al desarrollo constante de estas técnicas. En las últimas décadas, en la rama de la Inteligencia Artificial las meta-heurísticas poblacionales representan una poderosa herramienta para la solución de problemas complejos. El objetivo de este trabajo es exponer los resultados obtenidos a partir de la aplicación de técnicas paralelas utilizando la arquitectura CUDA a la meta-heurística poblacional “Optimización en Mallas Variables” (VMO), con el propósito de lograr una mejora en el costo computacional en la misma. En el trabajo se implementan en paralelo las etapas más costosas del algoritmo en el proceso de expansión y contracción de la malla para la obtención de la nueva población inicial. Los resultados obtenidos fueron validados utilizando una versión secuencial de la metaheurística.
Los tiempos de ejecución de ambos algoritmos se comparan, mostrándose mejoras
sustanciales en la versión paralela con el uso de CUDA.
In the field of optimization, both academic and industrial problems that occur daily are often complicated and NP-hard. The computational solution of such problems cannot be carried out in a precise manner in a reasonable period of time, and the use of hardware resources of computation for this solution greatly increases. To address this issue, the design of the solution methods should be based on the set of the advanced approaches of combinatorial optimization using the methods of parallelism and engineering scale. So there are different branches of research dedicated to the ongoing development of these techniques. In recent decades, in the field of Artificial Intelligence, population based meta-heuristics represent a powerful tool for solving complex problems. The objective of this paper is to present the results obtained from the application of parallel techniques using the CUDA architecture to the meta-heuristic "Variable Mesh Optimization" (VMO), with the purpose of achieving an improvement in the computational cost. In the work are implemented in parallel the more expensive stages of the algorithm in the process of expansion and contraction of the mesh for obtaining the new initial population. The results obtained were validated using a sequential version of the meta-heuristics. The runtimes of both algorithms are compared, showing substantial improvements in the parallel version with the use of CUDA.
In the field of optimization, both academic and industrial problems that occur daily are often complicated and NP-hard. The computational solution of such problems cannot be carried out in a precise manner in a reasonable period of time, and the use of hardware resources of computation for this solution greatly increases. To address this issue, the design of the solution methods should be based on the set of the advanced approaches of combinatorial optimization using the methods of parallelism and engineering scale. So there are different branches of research dedicated to the ongoing development of these techniques. In recent decades, in the field of Artificial Intelligence, population based meta-heuristics represent a powerful tool for solving complex problems. The objective of this paper is to present the results obtained from the application of parallel techniques using the CUDA architecture to the meta-heuristic "Variable Mesh Optimization" (VMO), with the purpose of achieving an improvement in the computational cost. In the work are implemented in parallel the more expensive stages of the algorithm in the process of expansion and contraction of the mesh for obtaining the new initial population. The results obtained were validated using a sequential version of the meta-heuristics. The runtimes of both algorithms are compared, showing substantial improvements in the parallel version with the use of CUDA.
Descripción
Palabras clave
Meta-Heurísticas, Optimización Basada en Mallas Variables (VMO), Propósito General sobre Unidades de Procesamiento Gráfico (GPGPU), Arquitectura Unificada de Dispositivos de Cómputo (CUDA), Meta-Heuristic, VMO, GPGPU, CUDA