I'm using SOLVER to change three variables in order to get a minimum
sum squared error. I've noticed that if I change the initial values
for the 3 variables before running solver, that the solver will return
different values! There should only be one value for each associated
with a minimum sum of squared error, and I have placed non constraints
on what Excel can change the three variables in order to reach that
minimum value for the target cell (e.g. the sum squared error value).
Therefore, what the values for those three cells are prior to running
the solver shouldn't matter, even if they say nothing at all. This
concerns me about the accuracy of SOLVER. Has anyone else had this
issue? If so, is there an agreed upon way to set up the initial
variables prior to running SOLVER?

Please let me know if I need to elaborate.


Thanks,
Hosley

"Augabog" wrote...
....
I'm using SOLVER to change three variables in order to get a
minimum sum squared error. I've noticed that if I change the
initial values for the 3 variables before running solver,
that the solver will return different values! There should
only be one value for each associated with a minimum sum of
squared error, and I have placed non constraints on what
Excel can change the three variables in order to reach that
minimum value for the target cell . . . Therefore, what the
values for those three cells are prior to running the
solver shouldn't matter, . . .

Well, you're dead wrong.

First off, Solver, like any other iterative numerical optimization
software, only iterates until the change between a specified number of
iterations is less than a specified threshold. If the initial values
lead to a plateau or local minimum, that's what Solver is going to
give as its result, even if it's not a GLOBAL minimum. There are
settings in Solver's Options dialog which YOU can change to increase
the odds of Solver winding up with a global minimum.

There's a lot of literature on the pitfalls of numerical optimization.
You might want to consider buying a textbook on the subject.