Cannot calculate PK-Parameters of a population when one individual simulation failed

[msevestre]

@Yuri05 @PavelBal
I believe this is a major issue and we should come up with a way to deal with this

Scenario:

• I run a simulation of 4000 Individuals. The simulation runs literally takes a couple of hours
• One simulation out of 4000 (for individual Id # 1521) crashes
• The exported CSV results do not have any entry for 1521
• Calculating the PK Analyses for this pop simulation crashes because of index mismatch

We had this discussion for the RE. I think we should fix this as it can lead to a lot of frustration

[PavelBal]

I would be ok with just ignoring these values.

[msevestre]

Ignoring is fine. Right now it crashes

[Yuri05]

We already discussed this previously (Open-Systems-Pharmacology/OSPSuite.ReportingEngine#284) and proposed to define an failure_threshold (max. acceptable percentage of failed individuals per population).
If the number of failed simulations is below this threshold - just ignore the crashed simulations. If above: throw an exception and cancel the whole run

[msevestre]

This is not what I am talking about, or at least it's not obvious to me that it is the same
The pop sim runs ok. I get 3999 results for 4000 individual. However calculating PK crashes because it expects 4000 results. We need to support results files that are not complete when we calculate pk parameters

[PavelBal]

I agree with Michael on this. Threshold is a good idea but should be part of getResults or whatever we have there. If results are already there, the PK parameter should be calculated.

[Yuri05]

ok, I see

[msevestre]

and when it takes many hours to run... THAT SUCKS :)

[PavelBal]

@msevestre This cannot be solved in R, right? I am having the same issue now... Do you have some experience how to deal with it?

[msevestre]

What I did was identified the individual and swap it with another one. If you have multiple like this, or if some stuff is random...then it will be more complicated

This is such a pain. I think we should fix this asap. I'll tackle this after relexp redesign

[PavelBal]

What I did was identified the individual and swap it with another one

How did you do that? calculatePKAnalyses takes on object of SimulationResults, which then has a mismatch of ID with the index of the individual. How can you swap it there?

[msevestre]

I swap at the population level
e.g. I identified the individual crashing
removed it from the pop
and ran the stuff again

[msevestre]

in hindsight: I would identify the wrong individual: for example 354
In the pop, copy 354 with 353
in the results copy 353 results to 354
then calculate the PK. If you have a big enough pop, having 2 identical individuals will have no effect on stats etc.

[msevestre]

I think that I have it implemented. The PK Values will return NAN for an individual that was not calculated thus insuring that the array will have consistent length