MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
107<br />
To better understand the effect that n has on the optimization observe first that if ν ≥ 1 then the<br />
second and third conditions in (5.75) are redundant and hence all values of n greater than 1 will<br />
have the same effect on the optimization. One should then work solely with values ranging<br />
between 0≤ν<br />
≤ 1.<br />
Interestingly, a number of properties can be derived:<br />
(*)<br />
• n is an upper bound on the fraction of error (i.e. the proportion of datapoints with ξ > 0),<br />
• n is a lower bound on the fraction of support vectors.<br />
These two properties are illustrated in Figure 5-16.<br />
In summary, n-SVR is advantageous over e-SVM in that it allows one to automatically adjust the<br />
sensitivity of the algorithm through the automatic computation of e. To some extent, this is<br />
equivalent to fitting a model of the noise on the data (assuming a uniform noise model). This is<br />
illustrated in Figure 5-15.<br />
Figure 5-15: Effect of the automatic adaptation of e using n-SVR. (Top) Data with no noise. (Bottom) Same<br />
dataset with a white noise. In both plots, n-SVR was fitted with C=100, n=0.05, and a Gaussian kernel with<br />
kernel width=0.021.<br />
© A.G.Billard 2004 – Last Update March 2011