Abstract: With the maturity of cloud computing,increasing number of data analysis tasks are executed in cloud.Machine learning algorithm is a necessary part of data analysis,and the hyperparameter optimization for machine learning algorithms is a time-consuming and resource-consuming process.However,little research focuses on the final cost for executing hyperparameter optimization.For this reason,the method of executing the hyperparameter optimization as quick as possible with the lowest cost was studied on the basis of not changing the performance of learning model such as accuracy,precision,recall rate.An optimized workflow instance model was generated,which consisted of multiple parallel branches and each branch sequentially executed multiple models on a server.Based on the dual-colored bin packing algorithm,the branches were organized in such a way that they had a similar execution time and can be completed almost at the same time.Experimental results demonstrated that the proposed approach could meet the deadline and reduce the cost at the same time.

Key words: cloud computing, cloud workflow, execution optimization, hyperparameter optimization

摘要： 随着云计算技术的成熟,越来越多的数据分析任务被放在云计算平台中处理。而面向数据分析应用的机器学习算法的超参数优化是一个非常耗时且耗费资源的过程。超参数优化执行的成本开销是用户关注的一个重要因素之一。目前,针对超参数优化的研究大部分以学习模型性能为目标,考虑成本开销的研究工作较少。由此研究了基于当前的超参数优化方法,在不改变学习模型性能(如准确率、查准率、召回率等)的基础上,使得超参数优化执行尽可能快的同时成本开销尽可能低。首先,生成一个包含多个并行分支的超参数优化工作流,每个分支上的所有任务都运行在同一台服务器上。然后通过有色装箱算法来决策这些分支所包含的任务。实验结果表明所提算法可以在保证执行时间的前提下减少成本开销。

关键词: 云计算, 云工作流, 执行优化, 超参数优化