Best Practices for Variant Modeling

[etaylo6]

I am working on a variant model to leverage CH for model selection. A large part of the utility of CH, in my view, is the ability to store variant models and select between them. We have identified two major types in the paper.

Syntactic Variants: Represent a rearrangement of the same functional relationship. Hyperedges have different codomains/domains, but the overall set of the two is the same.
Semantic Variants: A competing representation within the model.

However, one issue with implementing multiple syntactic variants is that they create a cycle if they are fully invertible.

hg.add_edge([E, V], G, rel=modulus_G, label='E,V->G')
hg.add_edge([G, V], E, rel=modulus_E, label='G,V->E')
hg.add_edge([E, G], V, rel=modulus_V, label='E,G->V')

If two of these variables are provided, we can solve for the 3rd and exit the pathfinding loop. However, for all other cases, we might get stuck in the loop.

There are certainly workarounds for preventing this behavior; the question is, what is the best practice?

My first try is an "existence" viability condition that works a priori. This edge does not exist if the input space does not have certain elements. This essentially means that "this relationship can never be calculated given the set of information provided."

[jmorris335]

One messy solution is to implement #48, which allows you to turn off cycles for the entire graph. That might work in some cases, but obviously not if you want cycles in other instances. A more granular solution was suggested in #1, but not implemented, which was to add a parameter marking certain nodes as not resolveable, so that the solver could never solve a cyclical index on them. (They could still be used in a cycle, but not individually resolved). A final option is to make cycles off by default, with the user having to specify a cycle that should be unraveled during solution.