Struct loopybayesnet::BayesNet

pub struct BayesNet { /* fields omitted */ }

Representation of a Bayesian Network

Once built by adding the nodes one by one, you can use it for inference computation on the graph given some evidence.

Methods

impl BayesNet

pub fn new() -> BayesNet

Create a new empty Bayesian Network

pub fn add_node_from_probabilities<D: Dimension + RemoveAxis>(
    &mut self,
    parents: &[usize],
    probabilities: Array<f32, D>
) -> usize

Add a new node to the network

You need to specify the list of its parents, and an array of probabilities representing p(x | parents). If the parents are (p1, ... pk), the shape of the array should thus be: (N, N_p1, ... N_pk), where N is the number of possible values for the current variables, and N_pi is the number of values of parent pi.

If the node has no parents, the propabilities must be single-dimenstionnal and represents a prior.

All values of probabilities should be finite, but the probabilities array does not need to be normalized, as it will be during the construction process.

pub fn add_node_from_log_probabilities<D: Dimension + RemoveAxis>(
    &mut self,
    parents: &[usize],
    log_probabilities: Array<f32, D>
) -> usize

Add a new node to the network from log-probabilities

Same as add_node_from_probabilities, but the input is in the form of log-probabilities, for greated precision.

All values of log-probas should be strictly smaller than +inf. -inf is valid and represents a probability of 0. The probabilities array does not need to be normalized, as it will be during the construction process. For example, the log-vector [0.0, -inf] will represent a vector of probabilities of [1.0, 0.0].

Log-probabilities are intepreted as computed with the natural logarithm (base e).

pub fn set_evidence(&mut self, evidence: &[(usize, usize)])

Sets the evidence for the network

Input is interpreted as a list of (node_id, node_value). Out-of-range evidence is not checked, but will result into a probability of 0.

pub fn reset_state(&mut self)

Resets the internal state of the inference algorithm, to begin a new inference

pub fn beliefs(&self) -> Vec<LogProbVector>

Compute the current state belief of each node according to the current internal messages

pub fn step(&mut self)

Compute one step of the Loopy Belief Propagation Algorithm

The algorithm can be run for any number of steps. it is up to you to decide when to stop.

A classic stopping criterion is when the yielded beliefs stop significantly changing.