Home for reduced factor graphs (socket for connecting queries with model spec)

[papamarkou]

Since model representation, inference and learning are going to be split following the ML paradigm (with which I personally agree), one question is where the reduced factor graph will be generated. I imagine that the code for the original factor graph will be hosted in OpenPP as part of the model specification. On the other hand, the queries related to inference will be coded in external packages (MCMC or GLM for example). The rising question is how the external packages will interact with OpenPP to deduce the reduced factor graph, as described in the inference framework of the proposal.

I can see two options; one would be that each external package will provide the implementation for the generation of the reduced factor graph, alongside with its inference code. The second option is to make sure that OpenPP is equipped with generic algorithms for the creation of the reduced factor graph (this falls within the scope of graphical model algorithms that are meant to have separate semantics from inference) and with an accompanying query API, which will allow external packages to define their queries using a universal framework. I find the latter approach more reasonable.