
234
where α
i
are the Lagrange multipliers estimated by
maximizing
(5)
with respect to α
i
, subject to the constraints
and 0 ≤ α
i
≤ C. Using the Karush-
Kuhn-Tucker (KKT) complementarity condition,
can be computed.
In (5),
is the so-called
kernel function, depending on the nonlinear func-
tion ϕ. The choice of the kernel K implicitly deter-
mines both ϕ and the high-dimensional feature
space (the range of ϕ). Several choices are possible
for the definition of K. In our experiments, we use
Gaussian (radial basis function) kerne ...