2021 Oct 22

Knowledge Is Made in Curious Ways

Since I began writing ode2dsp (mentioned previously), it has been refactored a few times, and has achieved what it set out to do. Success! Maybe.

When ode2dsp was first conceived, it was meant to be a tool to prevent silly mistakes in creating models. Having built this, it was not difficult to add other features that leverage the existing data structures—namely, computing stability conditions. I expected this to enable developers to anticipate stability issues in generated models.

For stability conditions to be useful, they have to calculated as closely as possible with respect to the generated code. This meant ode2dsp couldn’t delegate Newton’s method to the target code any longer; it has to compute it itself, otherwise it can’t calculate the stability condition. Although I implemented stability analysis for the entire mathematical pipeline—from ODEs to FDEs to approximated FDEs that map one-to-one to target code—ultimately, when the target code is diverging, only the last part of the pipeline matters.

Stability conditions are most useful when they are simple, which is rare. Generally, they are difficult to visualize because they depend on many variables. Sometimes, SymPy is not able to devise a simpler form, and this doesn’t mean one doesn’t exist. Stability analysis may benefit from manual tinkering. But it is more realistic to assume the developer doesn’t want to tinker; that they will see a big, complicated expression and assume the model has an operating point-dependent instability, which is quite common. And they’ll confirm this not by trying to plot the stability condition, but by running the model and looking at the waveforms, just like in the 1960s.

The relationship between a stability condition and the upstream definitions is not straightforward. It is difficult to understand how to improve a stability condition; how to change something upstream of it so as to reduce it or make it unconditionally true. I don’t have that kind of insight, and would think it unwise to confront that kind of complexity head-first.

Then I got an idea: what if we abandon the pipeline—stop modeling ODEs—and write stable FDEs directly? What if we parameterize the set of stable linear FDEs, then “nonlinearize” them by making their parameters functions of state, so as to guarantee stability? I’ve started investigating this, and it’s too early to say whether it’ll pan out, but I feel optimistic.

There’s an insight here, which isn’t about math, but people: I wanted to specify the stability condition, not calculate it. Stability turned out to be a hard requirement once I couldn’t take it for granted. Having learned this about myself, the old code won’t be needed any longer, despite over two months of development. I thought I was building a tool, but really I was building knowledge.


Discuss this post by emailing my public inbox. Please note the etiquette guidelines.

© 2021 Karl Schultheisz — Lancaster, PA, USA