[Ipopt] A large optimization problem

[Andreas Waechter]

Hi Krish,

Did you run your code through a memory checker?  Is this first element 
in IRN being set in your code?

We have been using MUMPS for a few years now, and it seems to run very 
robustly.  However, Stefan in his posting seems to indicate that MUMPS 
crashes (but I doubt that this is a hard crash like a seg fault, or rather 
some problem like exceeding memory).  You could of course try to see what 
happens when you use MA27 as an experiment.

But I would carefully check if you are setting the values correctly in 
your method that are called by Ipopt to get the sparsity pattern.  Maybe 
for some reason you skip the first entry.  It might really be helpful to 
run your code with a derivative checker (like valgrind on windows) on a 
preferrably small instance of your problem (otherwise you have to wait 
forever) - if you do that, I would set the Ipopt print_level very high, so 
that it will print out the values of the Jacobian and Hessian as well, and 
if there is something uninitialized, valgrind would complain about before 
the problem is detected in MUMPS.

Regards,

Andreas


On Fri, 1 Aug 2008, Krish Krishnan wrote:

> Hi all:
>
> I have a rather large portfolio optimization problem.  I am trying to use IpOpt with Mumps (due to licensing issues).
> I first outline the steps to create the problem.
>
> I derive a class COpt from IpOpt.
> An instance of COpt is p_optdata
>
> I next create the two smart pointers as follows:
> SmartPtr<COpt> p_optdata;
> SmartPtr<TNLP> p_optdata_tnlp;
>
> p_optdata = new COpt();
> p_optdata_tnlp = GetRawPtr(p_optdata);
>
> I set up the problem as follows:
>
> void CBMOptDoc::optimize()
> {
>    // Create an instance of the IpoptApplication
>    SmartPtr<IpoptApplication> app = new IpoptApplication(false, false);
>    // Initialize the IpoptApplication and process the options
>    ApplicationReturnStatus status;
>    status = app->Initialize();
>    if (status != Solve_Succeeded)
>    {
>        AfxMessageBox("Error during initialization");
>        return;
>    }
>    status = app->OptimizeTNLP(p_optdata_tnlp);
>
>    if (status == Solve_Succeeded)
>    {
>        Index iter_count = app->Statistics()->IterationCount();
>        Number final_obj = app->Statistics()->FinalObjective();
>    }
> }
>
>
>
>
>
> bool COpt::get_nlp_info(Index& n, Index& m, Index& nnz_jac_g,
>                        Index& nnz_h_lag, IndexStyleEnum& index_style)
> {
>    int nassets = ml.m_nAssets;
>    int kvars = nRI + nInd;
>    n = 2 * nassets;
>
>    m = 4 + kvars;
>
>    // nonzeros in the jacobian,
>    nnz_jac_g = (7 + 2 * kvars) * nassets ;
>
>    // and nonzeros in the hessian of the lagrangian
>    nnz_h_lag = (2 * nassets + 1) * (nassets + 1);
>
>    // We use the C index style for row/col entries
>    index_style = C_STYLE;
>
>    return true;
> }
>
>
>
> Here nassets = 501, kvars = 22 so the number of non-zeros are fairly 
> large.  I provide the needed jacobians and hessians.
>
> When I get into app->OptimizeTNLP(p_optdata_tnlp), the application gets 
> as far as line 1022 of dmumps-part7.f and crashes.  The reason is that 
> the arrays IRN and ICN in the module are corrupted .  A partial look at 
> IRN shows that IRN(1) has garbage in it.
>
> -        IRN    {...}    INTEGER(4)
>        IRN(1)    -842150450    INTEGER(4)
>        IRN(2)    1    INTEGER(4)
>        IRN(3)    2    INTEGER(4)
>        IRN(4)    2    INTEGER(4)
>        IRN(5)    3    INTEGER(4)
>        IRN(6)    3    INTEGER(4)
>
> Similarly ICN is also corrupted at ICN(1).
>
> Any thoughts on where I could be going wrong?  Appreciate your help as I have put in a lot of effort over the last two months on this.I wouldn't have gotten this far without your feedback. If needed I can provide the jacobian and hessians.
>
> Thanks a lot.
>
> Krish
>
>
>
>