Generative Datalog + Continuous Distributions

Generative Datalog has been introduced as part of the *probabilistic
programming Datalog language" by Bárány, ten Cate, Kimelfeld, Olteanu
and Vagena, combining features from declarative and probabilistic
programming. The heads of Generative Datalog rules may contain random
variables following some parameterised distribution, the parameters of
which are instantiated using constants or variables from the rule body.
Intuitively, such a program induces a database-valued stochastic
process. Generative Datalog could serve various purposes, for example
stochastic modelling and probabilistic inference with relational
databases or just as a representation system for probabilistic databases.

In this talk we review the language and its semantics, starting with the
original work by Bárány et al. for discrete parameterised distributions.
We then present our solution for the support of continuous
distributions, a problem that was left open in the original work. We
explore the difficulties, and point out some pitfalls regarding such an
extension, and discuss various properties of the language we obtain.