Solution to HW2 #3d,e using ready made SMO2 from Intro to PR book:

http://cgi.di.uoa.gr/~stpatrec/welcome3d.html -pkg download

http://www.elsevierdirect.com/companion.jsp?ISBN=9780123744869 -pkg divided by chapters

Result is the same .5% error rate as Peter's solution, but training seems quite a bit faster and allows better param adjustments.

Questions? email John at jklemail@gmail.com

%read data

[X1400, tokens, Y1400] = readMatrix('data/MATRIX.TRAIN.1400');

X1400 = full(X1400);

[Xtest, tokens, Ytest] = readMatrix('data/MATRIX.TEST');

Xtest = full(Xtest);

%params

C=.1;  %.1 gave best result from trying .5, 1

tol=.01;

maxiter=10^6;

eps=10^(-6);

kernel='linear';

method=0;

X=X1400;

Y=Y1400;

N1 = size(X, 1);

N2 = size(Xtest, 1);

Y(Y==0) = -1;  %convert 0's to -1's for smo

%train

[a, b, w] = SMO2(X, Y', kernel,0,0,C,tol,maxiter,eps,method);

nSVs = sum(a>0)

marg = 2/sqrt(sum(w.^2))

%test

h = w*Xtest' + b;

h(h>0) = 1;

h(h<0) = 0;

errate = sum(h~=Ytest)/N2