Bayesian Belief Network
A Bayesian belief network (also known as a BayesianNetwork, a BeliefNetwork or a BayesNet) is a representation of dependence (and so also independence) of random variables. It is named after ThomasBayes.
A belief network can be described in terms of
- structure: a directed acyclic graph where the nodes are random variables and a domain for each random variable
- function: how each variable depends on its parents, such as using a table, a decision tree, a "noisy or", or some parametrized form
- number: the actual numbers needed to specify a conditional probability of each variable given its parents
Algorithms for Inference in Belief networks
The main exact algorithms are
- VariableElimination? which exploits the sparseness of the graph to answer single queries
- CliqueTreePropagation? which creates a secondary structure to give the posterior on each variable and allow for the incorporation of new evidence.
- RecursiveConditioning? that allows for the tradeoff between time and space
Learning Belief networks
Alternative Related Representations
- A DecisionNetwork? adds decision varibles and utility variables for making sequential decisions
- A CausalNetwork makes predictions of interventions