The dagger library for probabilistic programming in OCaml. Currently implements:
- Single-site Metropolis-Hastings, a.k.a. lightweight Metropolis-Hastings
- Incrementalized single-site MH, similar to that implemented in Hakaru10
- Sequential Monte-Carlo, with systematic and stratified resampling plus facilities for defining custom resampling strategies.
The main package is prbnmcn-dagger. Packages prbnmcn-dagger-gsl and prbnmcn-dagger-stats
provide distributions implemented respectively through the GSL (GPL-licensed) and
prbnmcn-stats (MIT-licensed).
Look no further for the documentation and the examples it contains.
Some more examples are made available in the examples subdirectory. For now, you'll find:
- an implementation of a 2d ising model and a toy study of its behaviour around its critical temperature
- a basic kalman filter example
- an experiment on forecasting wind power production using an ad-hoc Kalman filter
- an implementation of the algorithm described in Adaptive approximate Bayesian computation by Beaumont et al.
Contributions and issue reports are welcome. Development currently happen on https://gitlab.com/igarnier/monorepo/ but I can take care of cherry-picking pull requests submitted here.
The name dagger refers to two things:
- a good mathematical framework for giving a semantics to probabilistic programming is a certain dagger category of Markov kernels, see eg this or that paper; Bayesian inversion corresponds to a particular symmetry of a mathematical structure and this symmetry is denoted using the † symbol.
- the underlying representation of the probabilistic model when using the incrementalized backend is as a directed acyclic graph (ie a DAG, which sounds exactly like the French translation of dagger)