|
Hi,
Providing the data is impossible with the 60K limit and no attachments...
The AlgLib part of my application does nothing different than the demo code - see below.
'ir_svm' just creates the (symmetric) N-D Hessian and (N+1)-D constraints matrices etc.
The constraints are: all Xi > 0 and sigma(Xi * Yi, i=1..N) = 0,
where N, the number of training points, was at most 770 - but might need to be larger,
Xi, i=1..N, are the LaGarange multipliers - and the QP unknowns,
where Yi, i=1..N, are either 1 or -1 according to Ti being a 'true' or 'false' training point.
Scales were set uniformly for all N - as 0.1, 1, 10, 100, 1000 - with no appreciable effect.
No initial guess for Xi is available - so initialized with ones.
The QP having converged successfully, W, defining the the M-dimensional (58) hyperplane separating the 'true' training points from the 'false' ones can be calculated - and a scalar b such that
(1)
<w . Ti> - b should be >= 1 for 'true' M-dimensional Training points and
<w . Ti> - b should be <= -1 for 'false' Training points;
<w . Ti> here is the inner product between the two M-dimensional vectors.
The distance between the two sets of training points is 2 / ||W||.
The result of the QP yields small but non-zero distance between the two sets -
but the conditions (1) don't hold correctly - the training points for which the Lagrange multiplies came out identically zero full fill - not (1) - rather, in most cases, a weaker version
(2)
<w . Ti> - b should be >= f1 for 'true' Training points and
<w . Ti> - b should be <= -f2 for 'false' Training points;
where 0< f1 & f2 < 1
- and in isolated cases not even (2) holds...
int _tmain(int argc, _TCHAR* argv[])
{
try
{
double *pa, *pb, *pc, *ps, *px;
int *pct;
int N = ir_svm(&pa, &pb, &pc, &pct, &ps, &px);
real_2d_array a;
a.setcontent(N, N, pa);
real_1d_array b;
b.setcontent(N, pb);
real_1d_array s;
s.setcontent(N, ps);
real_2d_array c;
c.setcontent(N + 1, N + 1, pc);
integer_1d_array ct;
ct.setcontent(N + 1, pct);
real_1d_array x;
x.setcontent(N, px);
minqpstate state;
minqpreport rep;
// create solver, set quadratic/linear terms
minqpcreate(N, state);
minqpsetquadraticterm(state, a);
minqpsetlinearterm(state, b);
minqpsetlc(state, c, ct);
// NOTE: for convex problems you may try using minqpsetscaleautodiag()
// which automatically determines variable scales.
minqpsetscale(state, s);
//
// Solve problem with BLEIC-based QP solver.
//
// This solver is intended for problems with moderate (up to 50) number
// of general linear constraints and unlimited number of box constraints.
//
// Default stopping criteria are used.
//
//minqpsetalgobleic(state, 0.0, 0.0, 0.0, 0);
minqpsetalgodenseaul(state, 1.0e-15, 5.0e+4, 10);
minqpoptimize(state);
minqpresults(state, x, rep);
printf("inneriterationscount %u\n"
"ncholesky %u, nmv %u\n"
"outeriterationscount %u\n"
"terminationtype %d\n"
, rep.inneriterationscount, rep.ncholesky, rep.nmv, rep.outeriterationscount, rep.terminationtype);
}
catch(ap_error)
{
printf("Exception!\n");
getchar();
}
return 0;
}
|
Login
Register
FAQ
Search
