jonathf / chaospy

Chaospy - Toolbox for performing uncertainty quantification.
https://chaospy.readthedocs.io/
MIT License
443 stars 87 forks source link

Doing a PCE for simulations with non-uniform simulation times #429

Open oarcelus opened 2 months ago

oarcelus commented 2 months ago

The question is simple. Is there a way of doing a PCE of a simulation where the simulations resulting from the different parameter samples result in simulations with different times?

The simulated time may be 60s before some cutoff is reached for example, in other cases, this cutoff is reached at 70s. Is there a proper way to treat this with chaospy?

jonathf commented 2 months ago

Am I understanding the problem correctly that you have a simulator that converges, but is costly enough that you want early stopping? Your idea here is to incorporate this cutoff value into a PCE somehow as a parameter to determine how much you can get away with?

I think the answer is no, but I might have misunderstood your question.

oarcelus commented 2 months ago

Hi, sorry, I might have not explained myself correctly. My simulations have some cutoff values for some parameters. When the parameter value hit this cutoff the simulation stops because it is reaching a limit where the calculations diverges if I continue to apply a load to it (lets say the system is failing above a given parameter limit).

This means that different runs for different parameter combinations will result in model responses that take different times to reach this cutoff value. Thus, it is not clear to me whether PCE is built to handle this. I can always try to normalize the simulation time and I know that the model response will still be different, but I do not know if it is the best approach.

jonathf commented 2 months ago

Are you saying that the response generated by the simulator is inconsistent?

So the short answer is then no, that is not handled by itself. You will need olto normalize and/or align the data manually.