25.12.2013 Views

CRANFIELD UNIVERSITY Eleni Anthippi Chatzimichali ...

CRANFIELD UNIVERSITY Eleni Anthippi Chatzimichali ...

CRANFIELD UNIVERSITY Eleni Anthippi Chatzimichali ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

In this research, computationally demanding tasks such as model evaluation and<br />

optimisation, also eminently time-consuming techniques such as bootstrapping, would<br />

greatly benefit from the application of parallel programming. The analysis pipeline<br />

that was presented in Section 2.2.4 involves several “embarrassingly parallel”<br />

iterations as summarised in Figure 3-2. Since these tasks run independently of each<br />

other, a divide-and-conquer parallelisation has been indeed feasible. There are several<br />

communication mechanisms that allow computational tasks to run cooperatively in<br />

parallel across a single multicore machine. In this work, the Message Passing<br />

Interface (MPI) (Gabriel et al., 2004) via the Rmpi (Yu, 2011) interface is employed.<br />

Figure 3-2 Embarrassingly parallel problems in the analysis pipeline<br />

The figure demonstrates all the iterations running in sequence within the analysis pipeline. Each of<br />

these problems runs independently, thus it can be easily parallelised.<br />

3.2.2 Approximation Algorithms<br />

Based on the overall accuracies of Figure 2-6, the ensembles of RBF SVMs were<br />

noticeably the most promising classification method. The generalisation performance<br />

of a nonlinear (RBF) SVM is greatly dependent on a suitable selection of and .<br />

Therefore, a two-step grid-search approach was initially applied as a means of<br />

hyperparameter optimisation (see Section 2.2.4). A grid-search over an arbitrary range<br />

of parameter values is an intuitive but computationally intensive technique, to the<br />

point that even parallel programming proved to be inadequate. In addition, the<br />

61

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!