smoker.m

支持向量机工具箱

% SMOKER Sequential Minimal Optimization for SVM (L1).
% [Alpha,bias,nsv,trn_err,margin]=smoker(K,labels,C,eps,tol,Alpha,bias)
%
% SMOKER is the SMO which uses precomputed kernel matrix 
%  insted of kernel function evaluations in each step.
%  It is faster but needs more memory compared to smo.c.
%
%  To make executable file run 'mex smoker.c'.
%
% Obligatory input:
%  K [NxN] kernel matrix for N training patterns.
%  labels [1xN] labels of training patterns (1 - 1st, 2 - 2nd class ).
%  C [real] or [2 x real] one trade-off constant for both the classes
%    or two constants for the first and the second class.
%
% Optional input:
%  eps [real] tolerance of KKT-conditions fulfilment (default 0.001).
%  tol [real] minimal change of optimized Lagrangeians (default 0.001).
%  Alpha [1xN] initial values of optimized Lagrangeians. If not given
%    then SMO starts from Alpha = zeros(1,N).
%  bias [real] initial value of the threshold. If not given then SMO
%    starts from bias = 0.
%
% Mandatory outputs:
%  Alpha [1 x N] found Lagrangeian multipliers.
%  bias [real] found bias.
%
% Optional outputs:
%  nsv [real] number of Support Vectors (number of Alpha > ZERO_LIM).
%  trn_err [real] classification error on the training data.
%  margin [real] margin between classes and the found hyperplane.
%
% See also SMO, SVMCLASS, SVM.
%

% Statistical Pattern Recognition Toolbox, Vojtech Franc, Vaclav Hlavac
% (c) Czech Technical University Prague, http://cmp.felk.cvut.cz.
% 
% Modifications
% 1-Nov-2001, V.Franc
% 21-Oct-2001, V.Franc
% 16-October-2001, V.Franc
% 30-September-2001, V.Franc, comments.
% 26-September-2001, V. Franc, created