Can state variables and control variables be treated differently?

[erica ye]

I encounter local infeasible situation when solving NLP. (I also tried
mult-start alg to get feasible solution. Failed)
When I give the same starting point, fminconset in Matlab can get the
local optimal solution.
I am wondering if it is related with the variable types.
If AIMMS can treat state and control variables differently, the size
of problem may be reduced.


I even try to use slack variables in the critical constraint to make
sure the problem is feasible.
If AIMMS can distinguish the control variable, it definitely can find
the feasible solution to the modified problem.
However, AIMMS fails to find feasible solution.
(I believe AIMMS model is correct, because I use the solution obtained
by Matlab to check if it is a feasible solution to the problem for my
AIMMS model.)

Thanks in advance!

[Guido Diepen]

Hi Erica,

please keep in mind that with NLP problems, there can be a lot of differences in the different solvers. Certain NLP solvers perform well on one problem, while others perform better on other problems.

If you provide a full starting solution you got from Matlab when starting the solve in AIMMS, does the solver in AIMMS see this solution as a valid solution?

What exactly do you mean with the difference between state and control variables? 

Guido Diepen

AIMMS Specialist

[erica ye]

Thanks for your reply！
Yes, AIMMS sees that solution as a valid solution.
Control Variable is independent while the state variable dependents on
the control variable.
For example:
control variable: x
state variable: y
y must satisfy the balance constraint y = f(x).

I am wondering if it will make it easier for AIMMS to solve the NLP
with real part and image part to represent all complex variables.
Right now I use magnitude and degree (Polar format) to represent a
complex variable while constraints are formatted in real part and
image part.
As an expert, do you have suggestions about this?

[Marcel Hunting]

Hi Erica,

Maybe you can try giving variable y the definition f(x) and make variable y inline (by using the Property attribute), and remove the y=f(x) constraint. That basically means that every occurrence of y in your model will be substituted by f(x).

Another option is to switch on the nonlinear presolver (see: Option Tree – Solvers general – Nonlinear presolve).

Marcel Hunting
AIMMS Software Developer

[erica ye]

Hi, Marcel
I tried to use the definition of the variable instead of additional
constraints.
But it gave me errors for "Division by 0". So I have to use the
additional constraints to represent it.
I've set the nonlinear presolve ON. still infeasible.

Do you think using the real part and image part instead of polar form
would help?
Thanks!