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.
102<br />
M<br />
∂ L =<br />
*<br />
i<br />
w − ∑ ( αi<br />
− αi<br />
) φ( x ) = 0;<br />
∂w<br />
i=<br />
1<br />
(5.68)<br />
∂L<br />
C (*) (*)<br />
= − 0<br />
i( *)<br />
i i<br />
ξ M α − η =<br />
∂<br />
(5.69)<br />
Substituting in (5.66) yields the dual optimization under constraint problem:<br />
M<br />
⎧⎧ 1 * *<br />
i j<br />
⎪⎪ − ∑ ( αi − αi<br />
)( α<br />
j − α<br />
j) ⋅ k( x , x )<br />
⎪⎪ 2 i, j=<br />
1<br />
max { ⎨⎨<br />
M<br />
M<br />
*<br />
αα , ⎪⎪ −<br />
* i *<br />
ε ( αi<br />
+ αi<br />
) + y ( αi<br />
+ αi<br />
⎪⎪ ∑ ∑ )<br />
⎩⎩ i= 1 i=<br />
1<br />
M<br />
* * i ⎡⎡ C ⎤⎤<br />
subject to ∑( αi<br />
− αi<br />
) = 0 and αi<br />
, αi<br />
∈ ⎢⎢ 0,<br />
i=<br />
1<br />
M ⎥⎥<br />
⎣⎣ ⎦⎦<br />
(5.70)<br />
Solving for the above optimization problem will yield solutions for the Lagrange multipliers α * , α .<br />
These can then be used to construct our estimate of the best projection vector w by replacing in<br />
(5.68):<br />
M<br />
w= −<br />
i=<br />
1<br />
*<br />
i<br />
( αi<br />
αi<br />
) φ( x )<br />
∑ (5.71)<br />
To determine the value of the offset b , one must further solve for the Karush-Kuhn-Tucker (KKT)<br />
conditions:<br />
i<br />
i<br />
( i<br />
y w ( x ) b)<br />
i<br />
i<br />
y w ( x ) b<br />
α ε + ξ + − , φ − = 0,<br />
i<br />
(<br />
i<br />
)<br />
α ε + ξ − + , φ + = 0,<br />
* *<br />
i<br />
and<br />
⎛⎛<br />
⎜⎜<br />
⎝⎝<br />
⎛⎛<br />
⎜⎜<br />
⎝⎝<br />
C<br />
M<br />
C<br />
M<br />
⎞⎞<br />
− αi<br />
⎟⎟ξi<br />
= 0,<br />
⎠⎠<br />
* ⎞⎞ *<br />
− αi<br />
⎟⎟ξi<br />
= 0.<br />
⎠⎠<br />
(5.72)<br />
The above optimization problem can be solved using interior-point optimization, see (Scholkopf &<br />
Smola, 2002).<br />
It is worth spending a little bit of time looking at the geometrical implication of conditions (5.72).<br />
(*) C<br />
The last two conditions show that, when the Lagrange multipliers take valueα i<br />
= , the<br />
M<br />
corresponding slack variable (*)<br />
can take arbitrary value and hence the corresponding points can<br />
ξ<br />
i<br />
be anywhere, including outside the e-insensitive tube. Vice-versa, when<br />
have ξ (*) i<br />
= 0 . These become the support vectors.<br />
(*) ⎤⎤ C ⎡⎡<br />
αi<br />
∈ ⎥⎥ 0,<br />
M ⎢⎢<br />
⎦⎦ ⎣⎣ , we<br />
© A.G.Billard 2004 – Last Update March 2011