TR with Boundry Constraints

[David_Gross]

I was having trouble with the RCI TR Routines with boundry constraints in C andstarted working with the example given and was able to reproduce my problem with that. It appears that for certain objective functions, it ignores my upper constraint.

To illustrate, if I change n = 2 and m = 2 in the example "ex_nlsqp_bc.c"and set the objective function equal to:

// broken
f[0] = x[0] - x[1];
f[1] = x[0] - 5;

// Working
f[0] = x[1] - x[0];
f[1] = x[1] - 5;

with constraints:

/* set the initial guess */
x[0] = -10.0;
x[1] = -10.0;

/* set bounds */
LW[0] = -20.0;
LW[1] = -20.0;

UP[0] = -3.0;
UP[1] = -4.0;

the result will be 5 and 5 for the "broken" one and -4 and -4 for the "working" one. I don't understand why these would produce different results and further don't understand why it ignores my bounds. Am I using it incorrectly?

Michael

[Nikita_S_Intel]

Hi Michael,

We know about this issue. In general, the TR solvers are solvers for nonlinear optimization problems, but your objective function is linear. In some casesth theTR solvers may work incorrectly for problems with linear objective functions with boundary constraints. We've fixed this issue so TR solvers work correctly for such types of problems. This fix will be available in next update of MKL 10.2.

Thank you!

--Nikita

[Artem_V_Intel]

Hi Michael,

Problem you reported has been fixed in MKL 10.2 Update 3. It is available now at the Intel Registration Center.

Thanks,
Art

[dbacchus]

Quoting - Artem Vorobiev (Intel)

Hi Michael,

Problem you reported has been fixed in MKL 10.2 Update 3. It is available now at the Intel Registration Center.

Thanks,
Art

Update 3 is available?

UPDATE: It is available! Why no sticky? Anyways - thanks, will be checking now...

[barragan_villanueva_]

Quoting - dbacchus

Update 3 is available?

UPDATE: It is available! Why no sticky? Anyways - thanks, will be checking now...

Yes! Now you can find it there.