The PILGRIM (ProbabIListic GRaphIcal Models) platform aims at providing efficient software tools in order to define and learn different probabilistic graphical models (currently directed ones: Bayesian networks and Probabilistic Relational Models), and use them to make inference on data. It is composed of 3 different projects:
- PILGRIM General offers common features for the other projects and allow to: define and serialize Bayesian networks; connect to tabular datasets; compute common measures such as KL-divergence;
- PILGRIM Structure Learning offers structure learning algorithms for Bayesian networks (e.g. greedy search, MMHC), structure evaluations (e.g. BDeu, BIC, SHD), and other related statistical tests (e.g. mutual information);
- PILGRIM Relational offers many features for the definition and learning of Probabilistic Relational Models (PRM) and their extensions, as well as their instanciation into grounded models for inference.
PILGRIM is actively developped and maintained by members of the Data, User, Knowledge (DUKe) team of the CNRS LSN Laboratory, in Nantes, France, with the help of Ouest Valorisation, Technology transfer acceleration company. The project is lead by Pr. Philippe Leray.
The PILGRIM platform comes in two flavors: a GPL 3.0 version, distributed from this website (you can download the latest version here); and an extended version with extra proprietary code. If you are interested in the latter case, please contact us.
The PILGRIM project relies on and encapsulates the ProBT platform, developped by ProbaYes. ProBT allows to model, learn and make inference in Bayesian networks and PILGRIM reuses many of its classes defining basic concepts in probabilistic programming, while extending its capabilities in terms of Bayesian networks structure learning and in terms of relational datasets support.