Parameter Learning for Markov Random Fields with Highest Confidence First Estimation

We study the problem of learning parameters of a Markov Random Field (MRF) from observations and propose two new approaches suitable for use with Highest Confidence First (HCF) estimation. Both approaches involve estimating local joint probabilities from experience. In one approach the joint probabilities are converted to clique parameters of the Gibbs distribution so that the traditional HCF algorithm can be used. In the other approach the HCF algorithm is modified to run directly with the local probabilities of the MRF instead of the Gibbs distribution