[Coin-ipopt] IPOPT as an Excel solver plugin

John Pye john.pye at anu.edu.au
Thu Dec 20 01:04:20 EST 2007

Damien Hocking wrote:
> There's a few things we could try there.  The duh-simple option would be to
> assume a dense system, but that would limit you to around 1000 variables.
> It won't really hurt the sparse solvers at that size, but it's very
> inefficient.  We could do a couple of perturbation passes to identify the
> sparsity pattern.  That's not perfect either, you can't guarantee that
> you'll pick up every Jacobian entry, especially in something like Excel
> where roundoff is more prevalent.  You could rebuild the system and restart
> from the current point if you detect a change in the Jacobian structure;
> this isn't as efficient as a structural incidence matrix but it's a lot
> better than assuming it's dense.  This is where I'm thinking of going.
> I'd do finite differences as the rule, with an exception for analyticals if
> the user really wants to.  My goal here is to provide a small-scale NLP
> solver for Excel for straightforward work, I'm not interested in trying to
> compete with Matlab or Maple.  I see this as being useful for up to a few
> hundred or a thousand variables at the most.  The issue won't be the size of
> problem that can fit in Excel, it will be the evaluation speed.  Anyway,
> it's worth a try.  

Isn't the idea with these spreadsheet solvers to create the sparsity
pattern by examining the symbolic structure of the equations,
cell-by-cell? I had imagined that that must be how Excel Solver works...


More information about the Coin-ipopt mailing list