forum.alglib.net

I currently use BLEIC for minimization solution. I implement a case from MSDN example in the following link http://msdn.microsoft.com/en-us/library ... 87(v=vs.93).aspx

The following is my source code.
void function1_grad(const real_1d_array &x, double &func, real_1d_array &grad, void *ptr)
{
//
// this callback calculates f(x0,x1) = 20*x0 + 15*x1
// and its derivatives df/d0 and df/dx1
// set goal (20 * sa + 15 * vz);

func = 20 * x[0] + 15 * x[1];
grad[0] = 20;
grad[1] = 15;
}

int main(int argc, char **argv)
{

// using BLEIC optimizer.

//set initial point, wild guess
real_1d_array x = "[3000.0,4500.0]";

//set scale
real_1d_array s = "[10.0,10.0]";

//set boundry
// 0 <= vz <= 9000,
// 0 <= sa <= 6000);

real_1d_array bndl = "[+0.0,+0.0]";
real_1d_array bndu = "[+6000.0,+9000.0]";

//set linear constrain
// 0.3 * sa + 0.4 * vz >= 2000,
// 0.4 * sa + 0.2 * vz >= 1500,
// 0.2 * sa + 0.3 * vz >= 500);

real_2d_array c = "[[0.3,0.4,2000.0],[0.4,0.2,1500.0],[0.2,0.3,500.0]]";

//set >= (1 ), = (0), <= (-1)
integer_1d_array ct = "[1,1,1]";

minbleicstate state;
minbleicreport rep;

//
// These variables define stopping conditions for the underlying CG algorithm.
// They should be stringent enough in order to guarantee overall stability
// of the outer iterations.
//
// We use very simple condition (gradian base) - |g|<=epsg
//
double epsg = 0.00001;
double epsf = 0;
double epsx = 0;

//
// These variables define stopping conditions for the outer iterations:
// * epso controls convergence of outer iterations; algorithm will stop
// when difference between solutions of subsequent unconstrained problems
// will be less than 0.0001
// * epsi controls amount of infeasibility allowed in the final solution
//
double epso = 0.0001;
double epsi = 0.0001;

//
// Now we are ready to actually optimize something:
// * first we create optimizer
// * we add boundary constraints
// * we add linear constraints
// * we set scale
// * we tune stopping conditions
// * and, finally, optimize and obtain results...
//
minbleiccreate(x, state);
minbleicsetbc(state, bndl, bndu);
minbleicsetlc(state, c, ct);
minbleicsetscale(state,s);
minbleicsetinnercond(state, epsg, epsf, epsx);
minbleicsetoutercond(state, epso, epsi);
alglib::minbleicoptimize(state, function1_grad);
minbleicresults(state, x, rep);

//
// ...and evaluate these results
//
printf("%d\n", int(rep.terminationtype)); // EXPECTED: 4
printf("%s\n", x.tostring(2).c_str()); // EXPECTED: [2,4]
//Sleep(5000);
return 0;
}

My question is when I set different initial points I get different answers, sometimes return "NAN"
case 1: set initial point, real_1d_array x = "[3000.0,4500.0]", return correct answer [2000, 3500]
case 2: set real_1d_array x = "[1000.0,1000.0]", return [NAN, NAN]

What is the problem resulted from? and how to fix it?

Hello!

I working on your problem right now. It seems that part of optimizer which tries to find feasible point is working incorrectly on your set of constraints. It seems that the problem is triggered by the start from the infeasible point. It is not fully confirmed yet, but start from feasible point should solve issue. I plan to give more information on your issue tomorrow. You can be sure that it has top priority for us.

Thanks for kindly help.

And, the question would be how to define feasible point? In case 2 (x[0]=1000 x[1]=1000), both are within their own boundary. could be considered as feasible points?