How do I specify prior knowledge in EM

[Anders L Madsen]

The content of the conditional probability tables (CPTs) prior to running the EM algorithm can be used to supply prior knowledge on the CPT. Using prior knowledge will imply that the CPTs (after EM) have been estimated based on prior knowledge and the data, i.e. the data and prior knowledge have been fused.
Associated with each node in the Bayesian network is an experience table. The EM learning algorithm estimate the CPT for each node with an experience table. If a experience table contains only zeros, then the initial content of the CPT is not taken into account when estimating the corresponding CPT. There is one experience count for each parent configuration. The experience count specifies the weight of the initial content of the CPT.