Prior distribution generating lots of samples at the end points

[CooperGarth]

Hi,

The last few times i've plotted my prior distribution i'm getting lots of pilot points generating permability values which are at the tail end of the distribution (see below).

It this because IES is "rounding" the tail ends of the distribution based on the min and max (see the image below)?

Also would possible solutions be:

-Change the standard deviation in the param.unc file to be smaller?
-Make the range larger
-Chose a different distribution (like uniform) if the range is small?

Any advice/help/suggestions would be appreciated!

Thanks in advance

[briochh]

@CooperGarth for IES there is an additional option, ies_enforce_bounds False, to have PESTPP not enforce the bounds that are in the control file.

However, you are correct, it is worth thinking about what it means to have a prior -- which theoretically represents expert knowledge of the feasible parameter distribution -- extending well beyond your parameter bounds -- which represent some sort of feasible minimum and maximum parameter value. Having most of your prior at or outside your bounds is probably not ideal.

If your uncertainty description in your .unc file is what you want honoured (double check it!), then modifying the parameter bounds in the control file is probably your best course (and maybe not enforcing bounds). However, if your bounds represent some sort of physical hard line for parameter values then you may want to revisit that uncertainty description. Remember you can also build your own ensemble and pass to IES with the ies_parameter_ensemble <filename> additional option.

[CooperGarth]

Thanks @briochh that did help - it did confirm that the tails of the distribution were being truncated.

It also appears the covariance matrix seems to play a big part in the distributions which seems obvious now.

[neilmanewell]

Just an additional note. Traditional PEST is more forgiving when incorrectly large variances (in relation to the range) are provided as regularisation information. It normalises them via a scaling factor. This is not the case with IES - you have to be more careful. By default (e.g. GW vistas, example PPCOV templates) variance is often just log 1, which may be too large for certain parameters. This trips so many people up. A good starting point for covariance is std=(logmax-logmin)/4, which approximately spans the 95th confidence interval based on prior knowledge. Starting with these nice log distributions can only help with history matching - you really want to avoid the insensitivity/potential model instability this truncation introduces. Use mkppstat and ppcov_sva to generate a spatial covariance matrix where you can too (not just a list of standard deviations).

Just thinking - perhaps PESTPP could warn user when supplied variance vs enforced bound range is significantly mismatched?

[jtwhite79]

Well said Neil. The issue of realized values stacked at the bounds can cause algorithmic and other downstream issues...I think we could pretty easily add a runtime check once the prior par ensemble is loaded/drawn to compare the empirical stdev to the bounds-implied stdev - good call