Compare DB Diff Calculations

[opotowsky]

The DB difference calcs are all done like so: 2 * (X - Y) / (X + Y)

This has some pros and cons, some discussed in the DiffResults docstring:

armi/armi/bookkeeping/db/compareDB3.py

Lines 95 to 98 in 2be5eb9

	All differences are based on a weird type of relative difference, which uses the
	mean of the reference and source data elements as the normalization value:
	2*(C-E)/(C+E). This is somewhat strange, in that if the two are very different, the
	reported relative difference will be smaller than expected. It does have the useful

I want to start a discussion on this calculation so we can handle the following scenarios better.

1. Difference is large but shows as small fraction

This is noted as an issue in the docstring. This could be bad since someone could miss this easily, thinking something isn't impacting a parameter.

2. Bounds SHOULD be +/- 2 but some diff values are much larger

If a negative value becomes positive or vice versa, the diff will be much larger than 2 in magnitude. This is an edge case we should filter for since we sometimes allow negative values even though they are non-physical.

3. There is a lot of confusion around what a difference value represents.

Ultimately, I would like all "normal" diffs to show as a fraction or a percent, and all "edge cases" to be better indicated as such (i.e., src is 10 and ref is 0 will always give a diff of 2, but that isn't obvious upon seeing the number 2).

Maybe we decide to change the calculation, or maybe we just institute better filtering. I don't know. I am looking for opinions!

[opotowsky]

@onufer @jakehader @john-science @albeanth @mgjarrett

[albeanth]

I agree with point 3 is your list above. It's really difficult to get a gut feel for how different something is. I often get confused on why some params show max(abs(diff)), mean(abs(diff)), and mean(diff), while some only show one or two of those three metrics. I usually just use the compre.log to indicate that something changed and use it to point me in a direction. But it's not naturally intuitive to see what's in the compare.log with what would be seen in XTVIEW.

If we had some hard coded logic to handle a few cases, I think a standard relative percent might serve us well. A few of the cases that come to mind that we would need to consider:

1. param in one db that isn't in another
2. param goes to zero in a db
3. really sensitive parameters constantly showing huge relative diffs
4. params that are vectors (do we use a max of the relative difference? or a stricter norm such as a 2-norm of the relative difference?)

If we find that we end up hardcoding a bunch of conditionals, then it might not be the right approach. Not sure.

On the other hand, if we don't want to use the compare.log to show something more intuitive based on relative differences, than I think it's ok to keep it as is (or some variant thereof), but it would be really nice to have that docstring more formally documented in the docs. It's a lot easier to read something formally written up in the docs than remembering which method within a given class in a specific file (that might move) has the relevant docstring (recognizing that I should really heed my own words in regard to the axial expansion changer 😄).

Thoughts?

[opotowsky]

1. param in one db that isn't in another

This is handled in a way already that I think is OK. This is where you see the src has <param> not in ref or vice versa

2. param goes to zero in a db

This will be shown as a +/- 2. It's maybe OK how we have it now, people just need to know the math and what 2 means. It's not the most intuitive but I don't have a better suggestion.

3. really sensitive parameters constantly showing huge relative diffs

Yes indeed. Small important diffs in one place might be ignored in favor of large unimportant diffs elsewhere. One partial remedy is education on which params get large swings in values we would rather ignore. Another is to add some kind of larger tolerance to these diff calcs (a messy suggestion to be sure!). Could perhaps be handled with a parameter setting. I don't know.

4. params that are vectors (do we use a max of the relative difference? or a stricter norm such as a 2-norm of the relative difference?)

This is handled in compareDB3. The diffs calcs include the possibility for either a scalar or vector parameter:

armi/armi/bookkeeping/db/compareDB3.py

Lines 441 to 442 in 2be5eb9

	mean = (src[()] + ref[()]) / 2.0
	diff = (src[()] - ref[()]) / mean

and

armi/armi/bookkeeping/db/compareDB3.py

Lines 462 to 465 in 2be5eb9

	absDiff = numpy.abs(diff)
	mean = numpy.nanmean(diff)
	absMax = numpy.nanmax(absDiff)
	absMean = numpy.nanmean(absDiff)

[opotowsky]

. I often get confused on why some params show max(abs(diff)), mean(abs(diff)), and mean(diff), while some only show one or two of those three metrics

This is because if you give a tolerance of, e.g., 1%, the max value diff may be over that and the mean value may be under. Sometimes max and mean match, so you get both. Sometimes both mean and max are over 1%, so you get both.

[opotowsky]

If we find that we end up hardcoding a bunch of conditionals, then it might not be the right approach.

Exactly. I want to make more rules so the compare log is more useful, but also not have a giant hardcoded specialty set of rules.

[albeanth]

This will be shown as a +/- 2. It's maybe OK how we have it now, people just need to know the math and what 2 means. It's not the most intuitive but I don't have a better suggestion.

What if it got treated the same way as the missing params get treated? E.g., <param> in src = <value>, is 0.0 in ref and vice-versa. That is probably a less than ideal message format in practice, but is an example off the top of my head.

Another is to add some kind of larger tolerance to these diff calcs (a messy suggestion to be sure!). Could perhaps be handled with a parameter setting.

I like the idea of a parameter setting. Could be something like compare_diff_tol_percent and/or compare_diff_ignore; where the former has a default of 1% and the latter has a default of False, that way we retain the current behavior, but can add some flexibility. It might be a good idea to have the compare.log print a summary of params with non-default values so they don't fall into the black hole of git history.

This is handled in compareDB3. The diffs calcs include the possibility for either a scalar or vector parameter:

Yeah totally. The original point was if we were to change it, what would we change it to?

This is because if you give a tolerance of, e.g., 1%, the max value diff may be over that and the mean value may be under. Sometimes max and mean match, so you get both. Sometimes both mean and max are over 1%, so you get both.

Makes sense. Would be great to have this in the docs.

but also not have a giant hardcoded specialty set of rules.

How many rules would we expect? At what point does it become burdensome?

[opotowsky]

What if it got treated the same way as the missing params get treated? E.g., <param> in src = <value>, is 0.0 in ref and vice-versa. That is probably a less than ideal message format in practice, but is an example off the top of my head.

This is totally worthwhile to consider!

This is handled in compareDB3. The diffs calcs include the possibility for either a scalar or vector parameter:

Yeah totally. The original point was if we were to change it, what would we change it to?

Oh I think the simple calcs (mean/max of absolute differences) are OK here. I would field arguments for someone who felt otherwise, though. It's always a tradeoff!

---

I'm thinking we could also keep a reference page for "diff quirks" that we can use internally so we aren't stuck coding up some really clunky/specific stuff in the framework. I don't know about adding much to the ARMI docs besides the basic functionality because the important differences we are discussing are plugin/reactor specific.

[albeanth]

We could always just get something put together in a form that is a step in the right direction and then revisit with different choices/improvements later on. I think simplifying/improving clarity and a brief docs section would be a great first step and give us a better feel for what we want to sink our teeth into next. @opotowsky thoughts?

[opotowsky]

Ok so some documentation on this is definitely in order. I'm thinking in the docstring is best, then it renders here: https://terrapower.github.io/armi/.apidocs/armi.bookkeeping.db.compareDB3.html

Agree? Disagree?

Also, are there any changes we outlined above that you'd be comfortable implementing? I like the idea of instead of +/- 2 having a printout of src is 0 and ref is not a la the added/deleted parameter printout. (not that actual printout. the one you have above is what I'd go with) This will help distinguish new params being set/unset from actual diffs.

[albeanth]

That is a good place, sure. I think it would be well served in the user docs as well. Possibly a new subsection? Could be under User Docs > Accessing Data in ARMI > Comparing Two Databases. Once docstrings get to a certain length and complexity, I think it's better to just link a docs page in the docstring instead. Just my 2 cents though.

I'll make a separate linked issue for replacing the +/- 2 and then whoever gets to it first can assign themselves!

[john-science]

@opotowsky Is this a ticket, or a Discussion?

[opotowsky]

This should be a discussion. I probably started it before I was accustomed to discussions...

[andfranklin]

Yeah totally. The original point was if we were to change it, what would we change it to?

Oh I think the simple calcs (mean/max of absolute differences) are OK here. I would field arguments for someone who felt otherwise, though. It's always a tradeoff!

I'd like to cast my vote to remove mean(diff). I think mean(abs(diff)) and max(abs(diff)) are probably the most helpful.