A Julia package for imprecise enhanced Bayesian Networks. Current functionality includes:
- Node types
- Discrete and continuous nodes, with conditional probability tables or distributions known a priori
- Discrete and continuous functional nodes, whose tables are derived from the parents through UncertaintyQuantification.jl models
- Imprecision at credal and simulation level — interval probabilities and probability boxes
- Network types
- Bayesian networks
- Credal networks (imprecise)
- Enhanced Bayesian networks — discrete, continuous, and functional nodes side by side
- Reduction & reliability analysis
- Discretization of continuous nodes
- Evaluation of functional nodes as structural reliability problems (Monte Carlo, Subset Simulation, Line Sampling, …)
- Imprecise reliability by Double Loop and Random Slicing
- Precise inputs reduce to a Bayesian network, imprecise inputs to a credal network
- Inference
- Exact inference by variable elimination
- Credal inference with lower/upper posterior bounds
- Parameter learning
- Precise estimates by maximum likelihood or Dirichlet smoothing, learning a Bayesian network
- Imprecise estimates by the Imprecise Dirichlet Model or Quesenberry–Hurst confidence intervals, learning a credal network
- One estimator for the whole network, or one per node
- Incomplete data: Expectation–Maximization for precise estimators, bounds over every completion for imprecise ones
- Visualization
- Layered, top-down network plots that encode each node's kind, precision, and discretization