Re: AW: IMP and parallelisation

On 9/20/2016 9:47 AM, Dirk Garmann wrote: > Thank you Leonid, > We have tried RANMETHOD=P, which is an interesting possibility. > Unfortunately this does not solve the issue. We will further evaluate if the > information from all nodes is used for the population update. > Any further hints are highly welcome > > Best > Dirk > > -----Ursprüngliche Nachricht----- > Von: Leonid Gibiansky [mailto:[email protected]] > Gesendet: Montag, 19. September 2016 22:26 > An: Dirk Garmann; [email protected] > Betreff: Re: [NMusers] IMP and parallelisation > > It is a good idea to use RANMETHOD=P at estimation step; then the > results should be identical independently of the number of nodes and > computer load. > > Concerning specific behavior .. looks strange. I would try to start from > the initial values of the model with the lowest OF and see what happens. > > Thanks > Leonid > > On 9/19/2016 1:29 PM, Dirk Garmann wrote: > > > Dear nmusers. > > > > During a popPK analysis using the M3 method and IMP we observed an > > unexpected behavior and would be interested if anyone else observed the > > same and can provide guidance/explanations. > > > > The IMP produces "strange" results in cases requiring a parallelization. > > > > We observed a general (and strong) trend that with increasing number > > of nodes the OBF increases (!) which in my opinion is unexpected if the > > number of samples in MC is sufficiently large. > > > > The initial settings have been as follows: > > > > Parse Type 1 > > > > $EST METHOD=IMP INTERACTION LAPLACIAN EONLY=0 ISAMPLE=300 NITER=1000 > > CTYPE=3 NOABORT GRD=SN(1,2) NOTHETABOUNDTEST PRINT=1 > > > > $EST METHOD=IMP INTERACTION NOABORT GRD=SN(1,2) EONLY=1 ISAMPLE=3000 > > NITER=30 PRINT=1 > > > > With 1 node the OBF decreased to ~- 1400 > > > > Using 16 nodes the OBF stabilized at ~ 1000 > > > > In both cases the OBF does not fluctuate much after 100 interations > > (monitoring of EM step) and seems to be stable (no clear hint for a > > local minima). > > > > Interestingly the estimated residual error is higher using 1 node. With > > 16 nodes the variability seems to be shifted to the ETAS. > > > > This behavior might be a concern for a covariate analysis using IMP > > > > Our first assumption was that we need to increase iSAMPLE in the EM > > step, since a different seed might be used for each node. However even > > increasing ISAMPLE to 3000 in the first step did not change the results > > much. > > > > My guess is that it points in the direction of how population values are > > updated, but I am not an expert in the implementation of IMP in NONMEM > > > > We would be highly interested in any guidance and explanation. > > > > Many thanks in advance > > > > Dirk > > > > Freundliche Grüße / Best regards, > > > > Dirk Garmann > > > > Head Quantitative Pharmacology > > > > Bayer Pharma Aktiengesellschaft > > > > BPH-DD-CS-CP-QP, Quantitative Pharmacology > > > > Building 0431, 322 > > > > 51368 Leverkusen, Germany > > > > Tel: +49 202 365577 > > > > Fax: > > > > Mobile: +49 175 3109407 > > > > E-mail: [email protected]_ > > > > Web: _ http://www.bayer.com_ > > > > Vorstand: Dieter Weinand, Vorsitzender | Christoph Bertram > > > > Vorsitzender des Aufsichtsrats: Hartmut Klusik > > > > Sitz der Gesellschaft: Berlin | Amtsgericht Charlottenburg, HRB 283 B