View source for Antenna Beamforming using Support Vector Machines

== Abstract ==

Support Vector Machines (SVM) have improved generalization
performance over other classical optimization techniques. Here, we
introduce an SVM-based approach for linear array processing and beamforming. The development of a
modified cost function is presented and it is shown how it can be applied to the problem of linear beamforning.
Finally, comparison examples are included to  show the validity of the new  minimization approach.



== Introduction ==

Support Vector Machines (SVM) have shown clear advantage in
prediction, regression and estimation over classical approaches in a
wide range of applications due to its improved generalization
performance. We introduce here a framework to apply the SVM approach
to linear antenna array processing.

Antenna array signal processing involves complex signals, for which a
complex formulation of the SVM is needed. We introduce this
formulation by introducing the real and imaginary parts of the error
in the primal optimization and then proceeding as usual to solve a
complex valued constrained optimization problem. We formulate the
algorithm using a modified cost function which includes the
unavoidable numerical regularization for the optimization. The
adjustment of the parameters into this cost function leads to an
improved robustness of the method against thermal and multiuser
noise.

The resulting algorithm is a natural counterpart of the real valued
Support Vector Regressor, which can be immediately applied to array
signal processing.

We introduce the derived formulation as the optimizer for a linear
beamformer. Several experiments illustrate the advantage of SVM over
the minimum mean squared error (MMSE)  based algorithms due to its robustness against outliers and
improved generalization ability. The first experiments compare the
behavior of both algorithms in an environment in which interferent
signals are close to the desired ones, thus producing nongaussian
noise. The last experiment illustrates the improved generalization
ability of the SVM when small data sets are used for training, which
is common in communications applications.



== The Support Vector Approach and the Cost Function ==

Let <math>\mathbf{x}[n]</math> be spatially sampled data. A linear
beamformer can be written as

<center><math>d[n]=\mathbf{w}^T \mathbf{x}[n]+\epsilon[n]</math></center>

where <math>{\mathbf{x}}[n]</math> is the vector of <math>M</math> elements of the array output
and <math>\epsilon[n]</math> is the output error.

Coefficients <math>\mathbf{w}</math>
are usually estimated through the
minimization of a certain cost function on <math>\epsilon[n]</math> 

The SVM approach can be applied to the adjustment of this model. The
main idea of SVM is to obtain the solution which has the minimum
norm of <math>\mathbf{w}</math>. Due to the minimization of the weight vector
norm, the solution will be regularized in the sense of Thikonov
[1], improving the generalization performance.
The minimization has to be subject to the constraints

<center><math>d[n]-{\mathbf{w}}^T {\mathbf{x}}[n] \leq  \varepsilon + \xi_n</math></center>
<center><math>-d[n]+{\mathbf{w}}^T {\mathbf{x}}[n]\leq  \varepsilon + \xi'_n</math></center>
<center><math>\xi[n], \xi'[n] \geq 0 </math></center>

not to be trivial. <math>\xi_n</math> and <math>\xi'_n</math> are the "slack"
variables or losses. The optimization is intended to minimize a cost
function over these variables. The parameter <math>\varepsilon</math> is used to
allow those <math>\xi_n</math> or <math>\xi'_n</math> for which the error is less that
<math>\varepsilon</math> to be zero. This is equivalent to the minimization of
the so-called <math>\varepsilon</math>-insensitive or Vapnik Loss Function
[2], given by

<center><math>L_{\varepsilon}(\epsilon)=
\begin{cases}
|\epsilon|-\varepsilon &|e|>\varepsilon\\
0&|e|<\varepsilon
\end{cases}</math></center>

The functional to be minimized is then

<center><math>L_p=||{\mathbf{w}}||^2+C\sum_n L_\varepsilon({\xi_n},{\xi'_n})</math></center>

subject to <math>\xi_n, \xi'_n \geq 0</math> where <math>C</math> is the trade-off between
the minimization of the norm (to improve generalization ability) and
the minimization of the errors [2].

The optimization of the above constrained problem through Lagrange
multipliers <math>\alpha_i</math>, <math>\alpha'_i</math> leads to the dual formulation[3]

<center><math>L_d=-({\boldsymbol \alpha}-{\boldsymbol \alpha'})^T{\mathbf{R}}({\boldsymbol \alpha}-{\boldsymbol \alpha'})+({\boldsymbol
\alpha}-{\boldsymbol \alpha'})^T{\mathbf{y}}-({\boldsymbol
\alpha}+{\boldsymbol \alpha'}){\mathbf{1}}\varepsilon</math></center>

to be minimized with respect to <math>({\alpha}_i-{\alpha'}_i)</math>.

It involves the Gram matrix <math>\mathbf{R}</math> of the dot products of the
data vectors <math>\mathbf{x}[n]</math>. This matrix may be singular and thus
producing an ill-conditioned problem. To avoid this numerical
inconvenience, a small diagonal <math>\gamma {\mathbf{I}}</math> is added to the
matrix prior to the numerical optimization.

We present here a modified derivation of the SVM regressor which leads to a
more convenient equivalent cost function '''(Fig.
\ref{fig:robustcost})'''

<center><math>L_{R}(e)=
\begin{cases}
0&|e|<\varepsilon\\
\frac{1}{2\gamma}(|e|-\varepsilon)^2-\varepsilon &\varepsilon \leq |e| \leq \varepsilon+e_C\\
C(|e|-\varepsilon)-\frac{1}{2}\gamma C^2 & e_C
 \leq |e|
\end{cases}</math></center>

where <math>e_C=\varepsilon+\gamma C</math>

This cost function provides a functional that is  numerically regularized by
the matrix <math>\gamma I</math>.  As it can be seen, the cost
function is quadratic for the data which produce errors between
<math>\varepsilon</math> and <math>e_C</math>, and linear for errors above <math>e_C</math>. Thus,
one can adjust the parameter <math>e_C</math> to apply a quadratic cost for
the samples which are mainly affected by thermal noise (i.e., for which the
quadratic cost is Maximum Likelihood). The linear cost is then
applied to the samples which are outliers [4],
[5] Using a linear cost function, the contribution of the
outliers to the solution will not depend on its error value, but  only on
its sign, thus avoiding the bias that a quadratic cost function produces.

Finally, we generalize the derivation to the complex-valued case, as it is
necessary for array processing.