The current research work follows the traditional approach for applying machine learning methods. Total learning time includes time for data collection as well as model generation. A large amount of cost and time is involved in the first phase of our experiments where we tried to collect sufficient data to explore any temporal variability even at the day level. Further to that, we were interested to generate multiple learning models using different machine learning algorithms while keeping the same observational criteria for the purpose of evaluation. Each of the experiment had a data collection cost of$153.888 involving 8 virtual machines belonging to different series and price range on Amazon EC2 running 24 hours for 7 days. On the other hand, learning cost varies for SVR and MPR. MPR is a representative of inference based methods and is used to get a better understanding of the actual relationship between a response variable and predictors. This learning method requires human intervention at various stages, as explained in Section 3.4.2, therefore requires more learning time. This learning effort, however, is reduced to some extent with the help of R-markdown script which can generate and display most of the important finding as well as visual graphs in nearly 30-60 minutes (depending on different data size, number of methods, iterations and computational speed). Based on the code generated findings, users with a different range of expertise can take few hours to a couple of days to generate a reasonable prediction model. These models are advantageous in certain aspects: Get a robust knowledge about the underlying relationship of response and predictors to generate a concrete set of outcomes. Moreover, use that knowledge to feed into complex machine learning methods to enhance the level of understanding. SVR is a representative of complex models which do
This, however, requires a lot of training time to adjust its parameters and does not clearly describe a transparent relationship of response and predictors at the end. This research follows the same approach of extracting useful information (from MPR based models) and to use it within complex learning methods (SVR based models) as explained in Section 3.5. By following this approach the training time of generated models ranges from a second to a maximum of 1 minute considering 10-fold cross-validation. This evidently describes a trade-off between the level of understanding and learning time.
A common assumption in a traditional learning setting is that the test and training data set are drawn from the same distribution and if the distribution changes then the lengthy process of rebuilding the model starts from the first step. Furthermore, the model derived for one type of distributional base data might not produce effective results for a different distribution. The change in distribution could be due to different applications or different cloud providers or virtual machines. This may result in having to repeat the approach from scratch by data collection. This leads us to think about generating a learning model be trained to produce an equally effective result with different distributional data. We tried resolving this matter by creating a generic model (Section 4.3.2) which can work equally effective on representative applications, yet not tested on different cloud providers. Conducting such experiments is still time-consuming and requires a cost for data collection. At this point, further challenges come into view from the perspective of cost-effectiveness that give rise to questions such as:
1. How can the cost and time be reduced when applying the machine learning technique? 2. How can we make our solution viable across different applications and cloud providers?
Our first intuition to answer above questions leading us to think about re-usability of existing knowledge that has been generated while creating learning models for different applications.
Chapter
5
The Transfer Learning Setting
5.1
Introduction
Chapter 4 has demonstrated that machine learning can play a vital role in designing an intelligent decision support system. Moreover, it provided the traditional principle of generating application specific as well as generic models using two machine learning methods, i.e. polynomial regression and SVR. The generated models are able to capture application behaviour on different deployment setups in order to make application-driven decisions. The chapter also examined the efficiency of the learning techniques, recognising that machine learning can impose significant training overhead.
Chapter 5 investigates enhanced learning techniques in order to make our proposed decision support system more efficient in terms of cost and time thus addressing the third research goal as stated in Chapter 1 and recalled here.
“The development and evaluation of an efficient decision-making method integrated with the estab- lished decision support system to reduce the learning and decision-making cost and to making it more cost-effective for use in cloud brokers.”
In particular, this chapter introduces a novel two-mode transfer learning scheme leading to sub- stantial reduction in the training overhead. The chapter also details the fundamentals of transfer learning technique and methods of transferring knowledge across domains. Furthermore, it explains how the two-mode transfer learning scheme is used to enhance the capability of our decision support system to make it more cost-effective for multi-cloud brokers. This transfer learning aided decision
support system is evaluated using different applications and two public cloud providers, namely AWS and GCE.