Dear Olaf, I don't agree with your analysis. The results shown by Teiji are deterministic. He made 2 series of 3 runs and the 2 series are identical. Jean
>> But, after MATLAB was restarted, the same values of objective >> function can be obtained. >> 1st 8740.2 >> 2nd 8731.9 >> 3rd 8728.3 >> < MATLAB is shutdown and restarted > >> 4th 8740.2 >> 5th 8731.9 >> 6th 8728.3 -----Message d'origine----- De : bounce-121567438-75398...@list.cornell.edu [mailto:bounce-121567438-75398...@list.cornell.edu] De la part de Olaf Schenk Envoyé : mardi 30 mai 2017 11:57 À : MATPOWER discussion forum <matpowe...@list.cornell.edu> Objet : Re: Different results at repeated execution Dear Jean, I think that Teiji is referring to the problem that if you use the multiple-threaded version of KNITRO you will have different solutions even in the case of using the same initial starting point. Regards, Olaf On 30.05.2017 11:50, MAEGHT Jean wrote: > Dear Teiji, > Knitro is using interior point methods. > Interior point method are known to have bad properties regarding initial > point (=if you start an interior point solver with an initial point which is > already an optimal point, the solver will not figure it out and may even > behave worse than starting from a dummy initial point). > But still, many times if you have a good guess or a good initial point, the > solver may behave better. > When you run your series of OPFs, the result is used as initial point for the > next OPF. This may explain different behavior of the solver. > > Implicit question: why are optimal values different? > 1/ as OPF is a non convex problem, you only get a local minimum, so > maybe you don't have the same local minimum each time 2/ stopping criteria > are using tolerances; maybe within the tolerances you used, there are several > solutions. > > For fine tuning of Knitro, you can see how we did in this paper: > https://arxiv.org/abs/1603.01533 > page 4, section V.A. > > knitro matlab documentation: > https://www.artelys.com/tools/knitro_doc/3_referenceManual/knitromatla > bReference.html > > If you force V and theta to zero in your case before each run, you wshould > always get the same result. > > Best regards, > -- > Jean Maeght > RTE - R&D Division > > > > -----Message d'origine----- > De : bounce-121566163-75398...@list.cornell.edu > [mailto:bounce-121566163-75398...@list.cornell.edu] De la part de Olaf > Schenk Envoyé : lundi 29 mai 2017 09:51 À : MATPOWER discussion forum > <matpowe...@list.cornell.edu>; Drosos Kourounis <kouro...@usi.ch>; > Kardoš Juraj <juraj.kar...@usi.ch> Objet : Re: Different results at > repeated execution > > Hi Teiji, > > I suggest to use an optimizer that has parallel bitwise reproducible > functionality. > > KNITRO is not able to give you identical results, but the > IPOPT/PARDISO > 5.0 binaries offers this functionality. You can it use the binaries > under "Matpower Libraries" on > > http://www.pardiso-project.org/#download > > Best, > > Olaf Schenk > > > On 27.05.2017 13:12, Teiji Ponishi wrote: >> Hi all, >> >> I use MATPOWER 5.1, and knitro as a optimization solver. >> >> When I run the MATPOWER repeatedly, the values of objective function >> can be slightly reduced as follows: >> >> 1st 8740.2 >> 2nd 8731.9 >> 3rd 8728.3 >> >> But, after MATLAB was restarted, the same values of objective >> function can be obtained. >> >> 1st 8740.2 >> 2nd 8731.9 >> 3rd 8728.3 >> < MATLAB is shutdown and restarted > >> 4th 8740.2 >> 5th 8731.9 >> 6th 8728.3 >> >> How do I obtain same results of MATPOWER without restarting MATLAB ? >> >> Best, >> >> Teiji > -- > Prof. Dr. Olaf Schenk > Advanced Computing Laboratory > Institute of Computational Science > Università della Svizzera italiana ** Switzerland > Via Giuseppe Buffi 13 ** 6900 Lugano > Phone: +41 (0) 79 368 22 81 ** Fax.: +41 (0)58 666 45 36 > Email: olaf.sch...@usi.ch ** http://www.ics.inf.usi.ch > > > > > "Ce message est destiné exclusivement aux personnes ou entités auxquelles il > est adressé et peut contenir des informations privilégiées ou > confidentielles. Si vous avez reçu ce document par erreur, merci de nous > l'indiquer par retour, de ne pas le transmettre et de procéder à sa > destruction. > > This message is solely intended for the use of the individual or entity to > which it is addressed and may contain information that is privileged or > confidential. If you have received this communication by error, please notify > us immediately by electronic mail, do not disclose it and delete the original > message." -- Prof. Dr. Olaf Schenk Advanced Computing Laboratory Institute of Computational Science Università della Svizzera italiana ** Switzerland Via Giuseppe Buffi 13 ** 6900 Lugano Phone: +41 (0) 79 368 22 81 ** Fax.: +41 (0)58 666 45 36 Email: olaf.sch...@usi.ch ** http://www.ics.inf.usi.ch "Ce message est destiné exclusivement aux personnes ou entités auxquelles il est adressé et peut contenir des informations privilégiées ou confidentielles. Si vous avez reçu ce document par erreur, merci de nous l'indiquer par retour, de ne pas le transmettre et de procéder à sa destruction. This message is solely intended for the use of the individual or entity to which it is addressed and may contain information that is privileged or confidential. If you have received this communication by error, please notify us immediately by electronic mail, do not disclose it and delete the original message."
