A belief network gives a probability distribution over a set of random variables. We cannot always expect an expert to be able to provide an accurate model; often we want to learn a network from data.

Learning a belief network from data has many variants depending on how much prior information is known and how complete the data set is. In the simplest case, the structure is given, all the variables are observed in each example, and only the conditional probabilities of each variable given its parents must be learned. At the other extreme, the agent may not know the structure or even which variables exist, and there may be missing data, which cannot be assumed to be missing at random.