[Coin-ipopt] Heuristic for IPOPT

[Dahai Xu]

Hi Andreas,

Many thanks for your information. The reason of modifying the core code of 
IPOPT is that my problem can not be exactly and efficiently formulated with 
smooth functions (e.g. a=min(b, c, d), etc.), whereas get_warm_start_iterate 
and "save_warm_start_iterate" are not implemented yet in the code.

It took me a while to figure out how to modify the IpoptData within you 
excellent object oriented framework. I tried to fix some variables after 
some iterations at two locations

1) in intermediate_callback: I modify some variables of x() by accessing 
dynamic_cast<DenseVector*>(const_cast<Vector*>(GetRawPtr(ip_data->curr()->x()))->Values()
2) in IpoptData::set_delta(SmartPtr<IteratesVector>& delta), I change the 
corresponding directions (components of delta) to 0 to avoid changing of 
these fixed variables.

Generally it works fine, but some Access violation will happen during the 
process of program destruction (i.e. after finishing main()).

Could you please comment a little more about my method?

Thanks.

Yours sincerely
Dahai Xu
----- Original Message ----- 
From: "Andreas Waechter" <andreasw at watson.ibm.com>
To: "Dahai Xu" <dahaixu at cse.Buffalo.EDU>
Cc: "coin-ipopt mailing list" <coin-ipopt at list.coin-or.org>
Sent: Tuesday, May 16, 2006 3:42 PM
Subject: Re: [Coin-ipopt] Heuristic for IPOPT


> Hi Dahai,
>
>> I would like to modify the core code of IPOPT to implement some 
>> heuristic. For example, a variable x is 0.01 at some iteration and I want 
>> to FIX it to 0 afterwards (i.e. future iterations). So where are the best 
>> locations / functions to do that?
>
> I don't think it would really work to change the problem definition during 
> the optimization.  In particular, fixing a variable to a value changes the 
> dimension of the problem, and the data structure cannot be adapted.  In 
> addition, I'm not sure how useful it is to fix a variable during the 
> optimization, since you don't know yet if it is really going to 0.  The 
> purpose of the interior point approach is not to make decisions about 
> activities.
>
> However, if you still want to try something, you could simply interrupt 
> the optimization when you realize that you want to fix a number of 
> variables and restart the algorithm from that point, after changing the 
> problem statement (I'm assuming that you are using Ipopt from a C++ 
> program.)
>
> You could overload the method intermediate_callback of the TNLP class, 
> which is called once per iteration (around the time of the convergence 
> check).  You could monitor the variables there and decide when you want to 
> quick (and return false in that case).  It is a bit tricky to get the 
> information about the variables out of the Ipopt internal data structures 
> (it is stored in the IpoptData object).  After the optimization was 
> aborted by you (by returning false), you could use the 
> get_warm_start_iterate method of the TNLP class, to set the starting point 
> for ALL Ipopt-internal variables.  All this is not documented, so you 
> would have to dig into the code...
>
> Regards,
>
> Andreas
>