<div>Peter</div>
<div> </div>
<div>Look at SICON 06 paper of Martin Burger (<a href="http://dx.doi.org/10.1137/05062723X">http://dx.doi.org/10.1137/05062723X</a>), He used IPOPT to solve discretized form of structural topology optimization.</div>
<div> </div>
<div>In my experience, IPOPT is *very* robust for such problems, but is very expensive in contrast to e.g. OC or MMA.</div>
<div> </div>
<div>Cheers</div>
<div> </div>
<div>RT<br><br></div>
<div class="gmail_quote">On Wed, May 27, 2009 at 9:06 AM, Clausen Peter M. <span dir="ltr"><<a href="mailto:peter.clausen@fe-design.de">peter.clausen@fe-design.de</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="PADDING-LEFT: 1ex; MARGIN: 0px 0px 0px 0.8ex; BORDER-LEFT: #ccc 1px solid">Hello<br><br>I am working at FE-DESIGN GmbH (<a href="http://www.fe-design.com/" target="_blank">www.fe-design.com</a>) and our<br>
program TOSCA Structure is a structural optimization<br>program. We use the Method of Moving Asymptotes (MMA of<br>Krister Svanberg) as optimizer which in most cases does a<br>really good job. In some cases though it has its drawbacks<br>
(linear object function, constraints which become active<br>later in optimization) and we would like to test an<br>alternative.<br><br>Our problems may exceed 1 mio. design variables and we<br>usually only need around 30-50 iterations to converge. The<br>
primal solution is a FEM system solved by an external<br>program, the sensitivity analysis (1. order) is done by<br>ourselves. We have no 2. order information.<br><br>My questions:<br>1) How would you expect Ipopt to perform in a benchmark<br>
against MMA ? Has anybody got experiences?<br>2) We “warmstart” the optimization each iteration and<br>leave the optimization to calculate the new primary<br>solution. Thus we have changed the MMA so that we call it<br>with objective function, constraints and sensitivities of<br>
the actual iteration plus some saved information from<br>former iterations, e.g. the last asymptotes and other<br>values. Could this be easily done with Ipopt ?<br><br>We would prefer to use the fortran version, but<br>interfacing to C++ is nothing new to us. It would simply<br>
mean that Ipopt’s objects would be created and destroyed<br>each iteration (would this be a performance problem?).<br><br>Best regards<br><br>Peter M. Clausen, developer at FE-DESIGN GmbH<br><br>_______________________________________________<br>
<br> Peter M. Clausen<br><br> Email: <a href="mailto:peter.clausen@fe-design.de">peter.clausen@fe-design.de</a><br> Tel. : +49 721 96467 - 236<br>_______________________________________________<br><br> FE-DESIGN GmbH, Haid-und-Neu-Str. 7, 76131 Karlsruhe,<br>
Germany<br> Registration Court: D-Mannheim HRB 107382<br> Director: Dr. Juergen Sauter<br> Fax: +49 (0)721 96467-290<br> Web: <a href="http://www.fe-design.com/" target="_blank">www.fe-design.com</a><br>_______________________________________________<br>
<font color="#888888"><br>--<br>This email was Anti Virus checked by Astaro Security Gateway. <a href="http://www.astaro.com/" target="_blank">http://www.astaro.com</a><br><br>_______________________________________________<br>
Ipopt mailing list<br><a href="mailto:Ipopt@list.coin-or.org">Ipopt@list.coin-or.org</a><br><a href="http://list.coin-or.org/mailman/listinfo/ipopt" target="_blank">http://list.coin-or.org/mailman/listinfo/ipopt</a><br></font></blockquote>
</div><br>