Home > Research > Publications & Outputs > A transfer learning-aided decision support syst...

Associated organisational unit

Electronic data

  • 2017Samreenphd

    Final published version, 8 MB, PDF document

    Available under license: CC BY-NC-ND: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License

Text available via DOI:

View graph of relations

A transfer learning-aided decision support system for multi-cloud brokers

Research output: ThesisDoctoral Thesis

Published
  • Faiza Samreen
Close
Publication date2017
Number of pages214
QualificationPhD
Awarding Institution
Supervisors/Advisors
Publisher
  • Lancaster University
Original languageEnglish

Abstract

Decision-making in a cloud environment is a formidable task due to the proliferation of service offerings, pricing models, and technology standards. A customer entering the diverse cloud market is likely to be overwhelmed with a host of difficult choices in terms of service selection. This applies to all levels of service, but Infrastructure as a Service (IaaS) level is particularly important for the end user given the fact that IaaS provides more choices and control for application developers. In the IaaS domain, however, there is no straightforward method to compare virtual machine performance and, more generally cost/performance trade-offs, within or across cloud providers. A wrong decision can result in a financial loss as well as a reduced application performance. A cloud broker can help in resolving such issues by acting as an intermediary between the cloud provider and the cloud consumer – hence, serving as a decision support system for assisting the customer in the decision process.

In this thesis, we exploit machine learning for building an intelligent decision support system which assists customers in making application-driven decisions in a multi-cloud environment. The thesis examines a representative set of appropriate inference and prediction based learning techniques, that are essential for capturing application behaviour on different deployment setups, such as Polynomial Regression and Support Vector Regression (SVR). In addition, the thesis examines the efficiency of the learning techniques, recognising that machine learning can impose significant training overhead. The thesis also introduces a novel transfer learning aided technique, leading to substantial reduction in this overhead. By definition, transfer learning aims to solve the new problem faster or with a better solution by using the previously learned knowledge. Quantitatively, we observed a reduction of approximately 60% in the learning time and cost by transferring the existing knowledge about the application and cloud platform in order to learn a new prediction model for some other application or cloud provider. Intensive experimentation has been performed in this study for learning and evaluation of proposed decision support system. Explicitly, we have used three different representative applications over two cloud providers, namely Amazon and Google. Our proposed decision support system, enriched with transfer learning methods, is capable of generating decisions that are viable across different applications in a multi-cloud environment. Finally, we also discuss lessons learned in terms of architectural principles and techniques for intelligent multi-cloud brokerage.