Computational models are ubiquitous, and using them involves mainly two things: fitting/inferring parameters, and making predictions using model+parameters. Pro

Computational models are ubiquitous, and using them involves mainly two things: fitting/inferring parameters, and making predictions using model+parameters. Probabilistic programming languages (PPLs) aim to provide a rich set of tools to make both these actions easy, given a model. To really exploit this to the fullest, the framework needs to also be able to handle uncertain parameters (rather than just "best fit" values), which goes over into the regime of "hierarchical" models i.e. probabilistic models for parameters in a probabilistic model. If we were to delve into details a little further, being able to make predictions from a generative stochastic model would call for efficient samplers, and being able to infer parameters might call for good variational optimization schemes (and various other methods on the landscale of Bayesian inference).