面向机器学习任务的调度方法研究  

Study on Scheduling Method for Machine Learning Tasks

作　　者：孙景玉 石振国 SUN Jing-yu;SHI Zhen-guo(College of Information Science and Technology,Nantong University,Nantong 226000,China)

机构地区：[1]南通大学信息科学技术学院,江苏南通226000

出　　处：《软件导刊》2020年第5期9-13,共5页Software Guide

基　　金：国家自然科学基金项目(61670171)。

摘　　要：为了提高机器学习任务执行效率并实现资源与任务的最佳匹配,在传统调度问题理论基础上对调度概念进行拓展,提出一种新的问题解决方案。该解决方案包括基于任务数据相似性原理,对任务集进行特征属性提取,构建以调度算法资源准确率较高为评价目标的数学模型。在考虑资源和任务匹配程度的前提下设计一种基于改进的简化粒子群优化的模糊C均值聚类算法,根据任务聚类结果设计新的基于机器学习任务聚类的任务调度算法。实验结果表明,构建的数学模型在大多数情况下性能良好,优化的聚类算法调用算法准确率比传统方法约高0.3~0.8个百分点,能够有效提高任务调度有效性。In order to improve the efficiency of machine learning tasks and achieve the best match between resources and tasks,the scheduling concept is extended based on the theory of traditional scheduling problems,and a new problem solution is proposed.The so⁃lution includes extracting feature attributes from the task set based on the principle of task data similarity,and constructing a mathe⁃matical model with high accuracy of scheduling algorithm resources as the evaluation target.Besides,a fuzzy C-means clustering algo⁃rithm based on improved simplified particle swarm optimization is designed based on the improved matching of resources and tasks.A new task scheduling algorithm based on machine learning task clustering is designed according to the clustering results of tasks.The ex⁃perimental results show that the constructed mathematical model performs well in most cases,and the algorithmic accuracy of the opti⁃mized clustering algorithm is about 0.3 to 0.8 percentage points higher than the traditional method,which effectively improves the effec⁃tiveness of task scheduling.

关 键 词：机器学习 任务调度算法 特征属性 C均值聚类算法 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]