Bayesian networks are a popular knowledge representation method that involve making the dependencies between variables explicit.  Several researchers have described techniques for learning Bayesian networks with continuous variables. These techniques assume a normal distribution at each node where the mean is a linear function of predecessor values and the variance is fixed. The learning task here is to find an appropriate combination of the coefficients for the linear function and the variance. This talk describes a method for learning Bayesian networks with continuous variables where each node is treated as a conditional multivariate normal distribution with a co-variance matrix and a mean vector. The learning task here is to find the entries of the co-variance matrix and the mean vector. This technique makes the dependence between the shared predecessors of two different nodes explicit and makes it possible to learn in Bayesian networks that can include deterministic linear functions.
 

Return to the seminar schedule