Motivation

Parameter estimation is a common task in many engineering and physics-focused disciplines. Typically, a parameterized model is proposed to explain a certain phenomenon, and its parameters are calibrated using some experimental data sets. Oftentimes, these models are then used in a design-context, where they predict the phenomenon’s behavior for a situation that is not covered by experiments. However, the designer might not have been involved in obtaining the parameter estimates she is using to evaluate her design case. Maybe these estimates lead to peculiar results in the considered design case, and should therefore be checked on their validity. Maybe there has been made a mistake. This is typically the point where it gets tricky. How can the parameter estimates in question be reproduced? What was the underlying model, which assumptions have been made, was there data that was excluded from the fit? What about the uncertainty of the estimates? Questions like these are often difficult to answer, even if the person or the people who conducted the calibration is/are still available.

Probeye was intended to address these reproducibility problems by providing an interface for defining parameter estimation problems in a way that can be translated into an ontology-based problem definition. This kind of tool-independent definition, that includes the model structure (forward models, parameters, priors, etc.) as well as the used experimental data is based on rdf-triples and can be saved as the fingerprint of the problem.

Next to the problem definition, also the inference data, i.e., the data generated by a solver while solving the problem should be saved in an ontology-based manner. Think for example about the parameter samples generate during an MCMC run. In order to do so, probeye comes with an integrated interface to different inference engines such as emcee or dynesty.