www.pudn.com > SVM_SteveGunn.rar > svr.m, change:1998-08-21,size:3982b


function [nsv, beta, bias] = svr(X,Y,ker,C,loss,e)
%SVR Support Vector Regression
%
%  Usage: [nsv beta bias] = svr(X,Y,ker,C,loss,e)
%
%  Parameters: X      - Training inputs
%              Y      - Training targets
%              ker    - kernel function
%              C      - upper bound (non-separable case)
%              loss   - loss function
%              e      - insensitivity
%              nsv    - number of support vectors
%              beta   - Difference of Lagrange Multipliers
%              bias   - bias term
%
%  Author: Steve Gunn (srg@ecs.soton.ac.uk)


  if (nargin < 3 | nargin > 6) % check correct number of arguments
    help svr
  else

    fprintf('Support Vector Regressing ....\n')
    fprintf('______________________________\n')
    n = size(X,1);
    if (nargin<6) e=0.0;, end
    if (nargin<5) loss='eInsensitive';, end
    if (nargin<4) C=Inf;, end
    if (nargin<3) ker='linear';, end  

    % tolerance for Support Vector Detection
    epsilon = svtol(C);

    % Construct the Kernel matrix
    
    fprintf('Constructing ...\n');
    H = zeros(n,n);  
    for i=1:n
       for j=1:n
          H(i,j) = svkernel(ker,X(i,:),X(j,:));
       end
    end

    % Set up the parameters for the Optimisation problem
    switch lower(loss)
      case 'einsensitive',
        Hb = [H -H; -H H];
        c = [(e*ones(n,1) - Y); (e*ones(n,1) + Y)];  
        vlb = zeros(2*n,1);    % Set the bounds: alphas >= 0
        vub = C*ones(2*n,1);   %                 alphas <= C
        x0 = zeros(2*n,1);     % The starting point is [0 0 0   0]
        neqcstr = nobias(ker); % Set the number of equality constraints (1 or 0)  
        if neqcstr
          A = [ones(1,n) -ones(1,n)];, b = 0;     % Set the constraint Ax = b
        else
          A = [];, b = []; 
        end
      case 'quadratic',
        Hb = H + eye(n)/(2*C);
        c = -Y;
        vlb = -1e30*ones(n,1);   
        vub = 1e30*ones(n,1);    
        x0 = zeros(n,1);              % The starting point is [0 0 0   0]
        neqcstr = nobias(ker);        % Set the number of equality constraints (1 or 0)  
        if neqcstr
          A = ones(1,n);, b = 0;      % Set the constraint Ax = b
        else
          A = [];, b = []; 
        end
      otherwise, disp('Error: Unknown Loss Function\n');
    end

    % Add small amount of zero order regularisation to 
    % avoid problems when Hessian is badly conditioned. 
    % Rank is always less than or equal to n.
    % Note that adding to much reg will peturb solution

    Hb = Hb+1e-10*eye(size(Hb));
 
    % Solve the Optimisation Problem
    
    fprintf('Optimising ...\n');
    st = cputime;
    
    [alpha lambda how] = qp(Hb, c, A, b, vlb, vub, x0, neqcstr);

    fprintf('Execution time : %4.1f seconds\n',cputime - st);
    fprintf('Status : %s\n',how);

    switch lower(loss)
      case 'einsensitive',
        beta =  alpha(1:n) - alpha(n+1:2*n);
      case 'quadratic',
        beta = alpha;
    end
    fprintf('|w0|^2    : %f\n',beta'*H*beta);  
    fprintf('Sum beta : %f\n',sum(beta));
    
    % Compute the number of Support Vectors
    svi = find( abs(beta) > epsilon );
    nsv = length( svi );
    fprintf('Support Vectors : %d (%3.1f%%)\n',nsv,100*nsv/n);

    % Implicit bias, b0
    bias = 0;

    % Explicit bias, b0 
    if nobias(ker) ~= 0
      switch lower(loss)
        case 'einsensitive',
          % find bias from average of support vectors with interpolation error e
          % SVs with interpolation error e have alphas: 0 < alpha < C
          svii = find( abs(beta) > epsilon & abs(beta) < (C - epsilon));
          if length(svii) > 0
            bias = (1/length(svii))*sum(Y(svii) - e*sign(beta(svii)) - H(svii,svi)*beta(svi));
          else 
            fprintf('No support vectors with interpolation error e - cannot compute bias.\n');
            bias = (max(Y)+min(Y))/2;
          end
        case 'quadratic',
            bias = mean(Y - H*beta);
      end 
    end

  end