A Cost-Effective Deadline-Constrained Scheduling Strategy for a Hyperparameter Optimization Workflow for Machine Learning Algorithms
| Autor: | Yan Yao, Jian Cao, Zitai Ma |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
Hyperparameter
Resource scheduling Computer science business.industry 02 engineering and technology Machine learning computer.software_genre Execution time Scheduling (computing) Workflow Multiple Models 020204 information systems Hyperparameter optimization 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Artificial intelligence business computer Algorithm Model building |
| Zdroj: | Service-Oriented Computing ISBN: 9783030035952 ICSOC |
| DOI: | 10.1007/978-3-030-03596-9_62 |
| Popis: | As a method of data analysis that automates analytical model building, machine learning is becoming increasingly popular. In most machine learning algorithms, hyperparameter optimization or tuning is a necessary step. Unfortunately, the process of hyperparameter optimization is usually computationally expensive and time-consuming. Currently, machine learning is becoming a service so that cost and time should be considered when a machine learning service is provided. In this paper, we propose a scheduling approach to satisfy two contradictory targets, i.e., cost and time, when models corresponding to multiple settings of hyperparameters need to be tried. In this approach, the execution time of the model with specific settings of the hyperparameters can be predicted first. Then we generate an optimized workflow instance model, which consists of multiple parallel branches and each branch sequentially executes multiple models on a server. Based on the number partitioning algorithm, the branches are organized in such a way that they have a similar execution time and can be completed almost at the same time. Through experiments on different machine learning algorithms, it demonstrated that this approach meets the deadline and reduce the cost at the same time. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|