Solutions at parameter bounds

[markcampanelli]

First, Thanks @HugoMVale for providing this ODRPACK update, along with so many modern language bindings.

I am translating my application that uses scipy.odr to odrpack-python. I am comparing results. I have several model parameters that are bounded below by zero, where zero is a legitimate value that represents an "ideal device". For example, a zero resistance model parameter. I handled this in a rather hacky way in scipy.odr, because bounds were not an option.

A couple issues I've hit:
(1) I have a result that drives a model parameter to its boundary at zero, and with odrpack-python I get "info=60002, stopreason='Questionable results or fatal errors detected. See report and error message.', success=False". However, the fit parameters look quite reasonable compared to my solution using scipy.odr (where I first detect a negative value in an initial fit then fix that parameter to zero and re-fit.)
(2) It would be nice to be able to fix elements of beta0 to values on the boundary of bounds.

Thoughts?

[HugoMVale]

Hello Mark,
thanks for the feedback. I suggest you try the following:

• Run the problem with report="long" to get information about the error.
• Check pages 34 and 35 of the ODRPACK95 guide. In particular, at the bottom of page 35 it is mentioned:

Note that INFO = 60000 indicates an error possibly caused by incorrectly specified user input to ODRPACK95, and more commonly by a poor choice of scale or weights, or a discontinuity in the derivatives.

[markcampanelli]

Hi Hugo,

When I run the same problem without any bounds (and also no parameter fixing), one of the parameters (named Gp, a conductance parameter) that should be (physically) non-negative goes from a positive initial condition to a negative final value during the fit, but interestingly I do not get the 60002 error condition in this case. The solution converges just fine. I do not see any errors reported in the errfile for either case, and TBH I'm not sure what exactly to look for in the rptfile.

It seems odd that the unconstrained version seems to have no issue going through zero for Gp, but the bounded version has (numerical?) problems/instabilities when Gp is bounded below by zero. I'm sorry that I do not have any good intuition about the algorithm here, but I'm happy to follow some pointers in a further investigation. It would appear that the Gp=0 boundary triggers some other numerical issue that is otherwise avoided in the unbounded case.

Also, it has not escaped me that there must be some considerable model and/or measurement error such that the best unconstrained fit makes one of the model parameters non-physical :).

-Mark