PI with random start values - if a run fails with drawn start values, it should be tried out with new values set.

[PavelBal]

If a run with random start values fails because of non-solvable start values, a new set of values should be tried out. To aviod endless trials, the user should be able to define the maximal amount of start values sets per run (e.g. 100).

As an example, out of 100 created runs in my PI, only two are running, and the rest failed at creation.

[msevestre]

Could you explain what non-solvable start values mean? I don't understand the issue that you are facing

[PavelBal]

The observed behavior is that the first evaluation of the run(s) is not possible - "RHS not finite" or something similar - and no start error can be calculated. Therefore, the run stops before performing any optimization.

My interpretation of this is that the particular set of the chosen parameter values leads to numerical problems in solving the simulation.

[msevestre]

@PavelBal ok understood. This change in the workflow will be quite hard to implement as the runs are "run" in parallel so intercepting the behavior to add more runs would be tricky.

In your case I would suggest the following

• Is it possible to change the RHS equations so that they are not so sensitive to parameter values?
• Another option would be for you to use 100 random values instead of the desired e.g. 10. As you mentioned earlier, a run stops immediately if the start values lead to an invalid parameterization. You could stop the PI as soon as you have reached the desired number of successful runs?

[Yuri05]

I would not automatically replace runs with invalid start values:

• If there are only few failed runs, it should not be a problem
• If there are many failed runs (e.g. >5%): PI problem might be not well stated and you might think about e.g. changing parameter ranges. In this case investigation of problematic runs would help you to adjust your PI input