RE: RE: method=ITS, OPTIMIZATION NOT TESTED (?!)
Dear Pavel
I do agree with Bob about increasing the number of samples. I wonder
what initial variance for the PK parameters you started with.
What is strange in the objective function pattern with iterations is
that it first goes in the wrong direction (first 3 iterations), then
goes down and stabilizes with very small Monte Carlo error (iteration 11
to 17) and then again goes up.
In addition of increasing ISAMPLE to 1000, I would suggest starting over
with larger initial variances. The initial pattern that I see reminds me
cases where small initial variances lead to an initial sampling region
that was not appropriate. I am starting usually with initial variances
equal 0.3 (and 1000 samples).
In addition, you could add in the same $EST an additional line which
would use FOCE INTERACTION. NONMEM7 would take the final estimates
obtained using IMP as initial estimates and since you have already the
answer, I am pretty sure that FOCE with INTERACTION will do the job and
you will get a nice objective function value (what you are used to).
Note that since FOCE with INTERACTION optimize a linearized version of
the model, the final objective function will not match perfectly with
the IMP one.
Best Regards;
Serge guzy
President, CEO; POP-PHARM; Inc;
PS: Getting indeed your control stream will help a lot.
Quoted reply history
From: [email protected] [mailto:[email protected]]
On Behalf Of Bauer, Robert
Sent: Tuesday, October 27, 2009 8:18 AM
To: [email protected]
Cc: [email protected]
Subject: RE: RE: [NMusers] method=ITS, OPTIMIZATION NOT TESTED (?!)
Pavel:
There is not a way to select the lowest objective function. Your
variations in the objective function are quite large during IMP, and it
suggests one of two things:
1) You may need to increase ISAMPLE to 1000
2) You may not be linear mu modeling all of the theta parameters
3) Just to make sure, are you using the release version of 7.1.0, rather
than the beta version
If you wish you may want to send me your control stream file.
Robert J. Bauer, Ph.D.
Vice President, Pharmacometrics
ICON Development Solutions
Tel: (215) 616-6428
Mob: (925) 286-0769
Email: [email protected]
Web: www.icondevsolutions.com
________________________________
From: [email protected] [mailto:[email protected]]
Sent: Tuesday, October 27, 2009 9:23 AM
To: Bauer, Robert
Cc: [email protected]
Subject: Re: RE: [NMusers] method=ITS, OPTIMIZATION NOT TESTED (?!)
Hello Robert,
As you suggested, I changed the options. The job converged much faster,
but the objective function was "quasioptimal", i.e. the final objective
function was equal to 10264.899631336189 (see below) while a lower
objective function of 10222.499515685291 was observed during the
interations when the default options were applied. I used the
parameters resulting in the objective function of 10222.499515685291 as
a starting point and run method=1 with interaction. The starting
objective function was equal to 10226.5427056802. I find method=imp
very useful for validation of the results. Is there a way to force it
to select the best parameters and the objective function?
Importance Sampling
MONITORING OF SEARCH:
iteration 0 OBJ= 10231.982923426112
iteration 1 OBJ= 10299.863138426574
iteration 2 OBJ= 10364.772068665749
iteration 3 OBJ= 10308.739048014546
iteration 4 OBJ= 10257.381411982740
iteration 5 OBJ= 10265.245304914670
iteration 6 OBJ= 10251.006105651519
iteration 7 OBJ= 10236.430245954567
iteration 8 OBJ= 10234.353761873708
iteration 9 OBJ= 10242.798081619065
iteration 10 OBJ= 10235.389009976587
iteration 11 OBJ= 10229.486328373108
iteration 12 OBJ= 10226.968499926681
iteration 13 OBJ= 10227.283866001104
iteration 14 OBJ= 10229.010908067567
iteration 15 OBJ= 10229.094607380130
iteration 16 OBJ= 10230.010670330952
iteration 17 OBJ= 10230.204979821499
iteration 18 OBJ= 10264.365519362043
iteration 19 OBJ= 10298.266732255344
Convergence achieved: ending mode
Elapsed estimation time in seconds: 13157.00
Evaluating one more iteration for Variance assessment:
iteration 19 OBJ= 10264.899631336189
OPTIMIZATION COMPLETED
Elapsed covariance time in seconds: 15766.53
Thank you,
Pavel
----- Original Message -----
From: "Bauer, Robert"
Date: Sunday, October 25, 2009 11:56 pm
Subject: RE: [NMusers] method=ITS, OPTIMIZATION NOT TESTED (?!)
To: [email protected], [email protected]
> Pavel:
> The objective function progress looks good. You should expect some
> Monte Carlo fluctuations. You should also run more iterations
> (perhapsNITER=200), and set CTYPE=3, which turns on the
> termination tester. To
> resume where you left off, rename your new control stream file,
> and put
> in the following lines.
>
> $EST METHOD=CHAIN NSAMPLE=0 ISAMPLE=50
> FILE=my_old_control_stream_file.ext
> $EST METHOD=IMP NITER=200 CTYPE=3 FILE=my_new_control_stream_file.ext
>
> Make sure you are linear MU referencing to get the greatest
> efficiency.
>
>
> Robert J. Bauer, Ph.D.
> Vice President, Pharmacometrics
> ICON Development Solutions
>
> Tel: (215) 616-6428
> Mob: (925) 286-0769
> Email: [email protected]
> Web: www.icondevsolutions.com
>
>
>
>
>
>
>
>
> ________________________________
>
> From: [email protected] [mailto:owner-
> [email protected]]on Behalf Of [email protected]
> Sent: Saturday, October 24, 2009 8:03 PM
> To: [email protected]
> Subject: [NMusers] method=ITS, OPTIMIZATION NOT TESTED (?!)
>
>
>
> Hello NONMEM Team,
>
> I found method=imp useful when there are local maxima.
> Nevertheless, at
> the end of optimization it prints a message, which makes me feel
> somewhat uncomfortable: OPTIMIZATION NOT TESTED. Also, the final
> objective function is not always the lowest one. An example is below.
> How do we interpret the results in this case?
>
> THETAS THAT ARE SIGMA-LIKE:
> MONITORING OF SEARCH:
>
> iteration 0 OBJ= 10625.663135214874
> iteration 1 OBJ= 10601.188754983375
> iteration 2 OBJ= 10537.895114114934
> iteration 3 OBJ= 10471.674625518765
> iteration 4 OBJ= 10430.297437731866
> iteration 5 OBJ= 10461.973668565577
> iteration 6 OBJ= 10462.638834406265
> iteration 7 OBJ= 10423.464983371641
> iteration 8 OBJ= 10417.959956991735
> iteration 9 OBJ= 10417.594007447198
> iteration 10 OBJ= 10414.708468642830
> iteration 11 OBJ= 10427.810693855947
> iteration 12 OBJ= 10412.889081059604
> iteration 13 OBJ= 10411.980622268416
> iteration 14 OBJ= 10424.501127174915
> iteration 15 OBJ= 10416.332869468861
> iteration 16 OBJ= 10416.622580251338
> iteration 17 OBJ= 10412.401585537709
> iteration 18 OBJ= 10415.117257355550
> iteration 19 OBJ= 10415.302370961055
> iteration 20 OBJ= 10409.066188189252
> iteration 21 OBJ= 10413.780620468329
> iteration 22 OBJ= 10410.787496174480
> iteration 23 OBJ= 10410.633582415931
> iteration 24 OBJ= 10409.970257443048
> iteration 25 OBJ= 10409.702420124611
> iteration 26 OBJ= 10409.213115058612
> iteration 27 OBJ= 10409.690639357370
> iteration 28 OBJ= 10410.016047785200
> iteration 29 OBJ= 10408.157468814226
> iteration 30 OBJ= 10407.779614704938
> iteration 31 OBJ= 10410.164563157052
> iteration 32 OBJ= 10408.364552302961
> iteration 33 OBJ= 10407.431920727997
> iteration 34 OBJ= 10408.286189641487
> iteration 35 OBJ= 10407.907347050501
> iteration 36 OBJ= 10407.451608770069
> iteration 37 OBJ= 10407.189482360372
> iteration 38 OBJ= 10406.484357336147
> iteration 39 OBJ= 10409.167125968375
> iteration 40 OBJ= 10406.840873883246
> iteration 41 OBJ= 10407.679485561714
> iteration 42 OBJ= 10405.341101045238
> iteration 43 OBJ= 10404.704382334516
> iteration 44 OBJ= 10405.348023082915
> iteration 45 OBJ= 10405.347406984720
> iteration 46 OBJ= 10401.873473651774
> iteration 47 OBJ= 10404.036204419035
> iteration 48 OBJ= 10405.072916975221
> iteration 49 OBJ= 10402.976628923887
> Elapsed estimation time in seconds: 30420.73
> iteration 50 OBJ= 10403.285958168881
>
> #TERM:
> OPTIMIZATION NOT TESTED
>
> Thanks,
>
> Pavel
>
>
>
>
>
>
>
>
>
>
>