基于GPU的SVM参数优化并行算法

唐美丽; 张劲松; 李璐; 马廷淮

doi:10.3969/j.issn.1671-7775.2017.05.013

PDF(1843 KB)

江苏大学学报（自然科学版） ›› 2017, Vol. 38 ›› Issue (5) : 576-581. DOI: 10.3969/j.issn.1671-7775.2017.05.013

论文

基于GPU的SVM参数优化并行算法

作者信息 +

Optimization algorithm of SVM parallel parameters based on GPU

Author information +

文章历史 +

摘要

为了缩短支持向量机(support vector machine,SVM)参数优化时长,提高SVM参数优化的效率,提出了基于图形处理单元(graphic processing unit,GPU)的SVM参数优化并行算法．分析了基于网格搜索和粒子群优化算法的并行特性,基于GPU设计了该优化算法的并行化方案,并在单GeForce GT 650M GPU卡上进行了试验验证．结果表明,并行化网格搜索和并行化粒子群参数优化算法不仅可以取得与非并行化参数优化算法相同的优化效果,而且执行时间大大减小,其中并行粒子群参数优化算法的加速比可高达2685,大幅提升了SVM的参数优化效率．

Abstract

To shorten the parameter optimization time of support vector machine(SVM) and improve the efficiency of parameter optimization, the optimization algorithm of SVM parallel parameters was proposed based on graphic processing unit (GPU). The parallel possibility was analyzed for grid search (GS) algorithm and particle swarm optimization (PSO) algorithm. Based on GPU, the parallel methods of GS and PSO were proposed. The experiments were executed on GeForce GT 650M GPU card to conduct 5cross validation. The results show that compared with the original optimization algorithms, the proposed parallel optimization algorithm can obtain the same optimum performance with greatly reduced execution time. The speedup ratio can reach 26.85 for parallel PSO algorithm, and the SVM parameter optimization performance is improved greatly.

导出引用

唐美丽, 张劲松, 李璐, 等. 基于GPU的SVM参数优化并行算法[J]. 江苏大学学报（自然科学版）, 2017, 38(5): 576-581 https://doi.org/10.3969/j.issn.1671-7775.2017.05.013

TANG Mei-Li, ZHANG Jin-Song, LI Lu, et al. Optimization algorithm of SVM parallel parameters based on GPU[J]. Journal of Jiangsu University(Natural Science Edition), 2017, 38(5): 576-581 https://doi.org/10.3969/j.issn.1671-7775.2017.05.013

参考文献

［1］CHAKRABARTI G, GROVER V, AARTS B, et al. CUDA: compiling and optimizing for a GPU platform ［J］. Procedia Computer Science, 2012, 9(9): 1910-1919.
［2］GRASSO I, PELLEGRINI S, COSENZA B, et al. A uniform approach for programming distributed heterogeneous computing systems［J］. Journal of Parallel and Distributed Computing, 2014, 74(12):3228-3239.
［3］SOPYA K, DROZDA P,GRECKI P. SVM with CUDA accelerated kernels for big sparse problems［C］∥Proceedings of the 11th International Conference on Artificial Intelligence and Soft Computing. Heidelberg: Springer Verlag, 2012: 439-447.
［4］HERREROLOPEZ S, WILLIAMS J R, SANCHEZ A. Parallel multiclass classification using SVMs on GPUs［C］∥Proceedings of the 3rd Workshop on GeneralPurpose Computation on Graphics Processing Units. New York: ACM,2010: 2-11.
［5］DALI N, BOUAMAMA S. GPUPSO: parallel particle swarm optimization approaches on graphical processing unit for constraint reasoning: case of maxCSPs［J］. Procedia Computer Science, 2015, 60: 1070-1080.
［6］CALAZAN R M, NEDJAH N, DE MACEDO M L. Parallel GPUbased implementation of high dimension particle swarm optimizations［C］∥Proceedings of the 2013 IEEE 4th Latin American Symposium on Circuits and Systems. Washington, D C: IEEE Computer Society, 2013,doi: 101109／LASCAS.20136518991.
［7］蔡勇,李光耀,王琥.基于CUDA的并行粒子群优化算法的设计与实现［J］.计算机应用研究, 2013, 30(8): 2415-2418.
CAI Y, LI G Y, WANG H. Design and implementation of parallel particle swarm optimization algorithm based on CUDA ［J］. Application Research of Computers, 2013, 30(8): 2415-2418. (in Chinese)
［8］WANG X G, LIU X, JAPKOWICZ N, et al. Ensemble of multiple kernel SVM Classifiers ［C］∥Proceedings of the 27th Canadian Conference on Artificial Intelligence. Heidelberg: Springer Verlag, 2014: 239-250.
［9］刘琰.支持向量机核函数的研究［D］.西安: 西安电子科技大学, 2012.
［10］RIVASPEREA P, COTARUIZ J, CHAPARRO D G, et al. Support vector machines for regression: a succinct review of largescale and linear programming formulations ［J］. International Journal of Intelligence Science, 2013, 3(1):5-14.
［11］YAN Z P, DENG C, ZHOU J J, et al. A novel twosubpopulation particle swarm optimization［C］∥Proceedings of the 10th World Congress on Intelligent Control and Automation. Piscataway: IEEE, 2012: 4113-4117.