欢迎来到天天文库
浏览记录
ID:5277354
大小:590.84 KB
页数:15页
时间:2017-12-07
《人工智能:基于GPU的约束网络模型和并行弧相容算法》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、计算机研究与发展DOI:10.7544桙issn1000‐1239.2017.20150912JournalofComputerResearchandDevelopment54(3):514528,2017基于GPU的约束网络模型和并行弧相容算法李哲李占山李颖(吉林大学计算机科学与技术学院长春130012)(符号计算与知识工程教育部重点实验室(吉林大学)长春130012)(zslizsli@163.com)AConstraintNetworkModelandParallelArcConsistencyAlgorithmsBased
2、onGPULiZhe,LiZhanshan,andLiYing(CollegeofComputerScienceandTechnology,JilinUniversity,Changchun130012)(KeyLaboratoryofSymbolicComputationandKnowledgeEngineering(JilinUniversity),MinistryofEducation,Changchun130012)AbstractConstraintsatisfactionproblemisapopularparadig
3、mtodealwithcombinatorialproblemsinartificialintelligence.Arcconsistency(AC)isoneofbasicsolutioncompressionalgorithmsofconstraintsatisfactionproblem,whichisalsoacorealgorithmofmanyexcellentadvancedalgorithms.Whenconstraintsareconsideredindependently,ACcorrespondstothes
4、trongestformoflocalreasoning.AneffectiveunderlyingACcanimprovetheefficiencyofreducingthesearchspace.Recentyears,GPUhasbeenusedforconstitutingmanysupercomputers,whichsolvemanyproblemsinparallel.BasedonGPU摧scomputation,thispaperproposesaconstraintnetworkspresentationmod
5、elN‐EanditsparallelgenerationalgorithmBuildNE.Accordingtofine‐grainedarcconsistencyAC4,aGPUGPUparalleleditionAC4anditsimprovedalgorithm—AC4+,areproposed.ThetwoparallelalgorithmsextendarcconsistencytoGPU.Experimentalresultsverifythefeasibilityofthesenewalgorithms.Compa
6、redwithAC4,theparallelversionshavemadethe10%to50%accelerationinsomesmallerinstances,andobtained1to2ordersofmagnitudeinsomebiggerinstances.Theyprovideacorealgorithmtootherconstraintsatisfactionproblemsolvinginparallelforfurtherstudy.Keywordsartificialintelligence;const
7、raintsatisfactionproblem(CSP);arcconsistency(AC);GPU;computeunifieddevicearchitecture(CUDA)摘要弧相容算法是约束满足问题的基本压缩求解空间算法之一,很多优秀的高级算法都以高性能的弧相容算法作为核心.近年来,以GPU为计算工具加速并行计算被用来尝试解决许多问题.基于GPU和基本的并行算法,提出一种适合GPU运算的约束网络表示模型N‐E,给出其生成算法BuildNE.结合GPUGPU细粒度的弧相容算法———AC4,基于N‐E模型提出AC4的并行化
8、算法AC4与改进算法AC4+,使弧相容算法得以扩展到GPU上执行.实验结果验证了该算法的可行性,与AC4算法的比较,其在一些规模较小的问题上取得了10%~50%的加速,在一些规模较大的问题上则加速1~2个数量级.为今后进一步在GPU上以并行形式解决
此文档下载收益归作者所有