Implementation and derivation of "Variational Bayesian inference for a nonlinear forward model." [Chappell et al. 2008] for arbitrary, user-defined model errors.
This motivation is certainly correct, but could be improved by an example to follow along. As @aklawonn wants to tackle this, his current wood compression tests (using dummy data) could be this example. Another idea of mine would be a concrete bridge where one model is the full bridge (1D beam...) with displacement/inclination sensors, another one is a compression test from multiple drilled, cylindrical core samples. Maybe, @aklawonn, you can update us on the general idea here in this issue.
Anyways, we should illustrate:
the advantages of named parameters
that not every parameter must be latent
how one (expensive forward) model can return several outputs in a structured way
that we can combine multiple model errors (aka experiments?)
that parameters of multiple experiments can be inferred simultaneously
the complexity of the example, e.g. by writing down the loglike for the problem (10+ terms! :smile: ) and sketching the laborious and error-prone process of changing things in there
interface to several solvers (VB algorithm, taralli, pymc3,...)
What else?
IMO the challenge will be to introduce the features above in a reasonable order. I have some kind of multi-step example in mind that increases in complexity. Starting with a simple loglike example that is familiar to the reader, introduce named parameters next, combine two examples and so on and so on.
This motivation is certainly correct, but could be improved by an example to follow along. As @aklawonn wants to tackle this, his current wood compression tests (using dummy data) could be this example. Another idea of mine would be a concrete bridge where one model is the full bridge (1D beam...) with displacement/inclination sensors, another one is a compression test from multiple drilled, cylindrical core samples. Maybe, @aklawonn, you can update us on the general idea here in this issue.
Anyways, we should illustrate:
What else?
IMO the challenge will be to introduce the features above in a reasonable order. I have some kind of multi-step example in mind that increases in complexity. Starting with a simple loglike example that is familiar to the reader, introduce named parameters next, combine two examples and so on and so on.