EM Learning of Product Distributions in a First-order Stochastic Logic Language

D. Pless and G. Luger (USA)


Stochastic Modeling, Knowledge Representation


We describe a new logic-based stochastic modeling language called Loopy Logic. It is an extension of the Bayesian logic programming approach of Kersting and De Raedt [1]. We specialize the Kersting and De Raedt formalism by suggesting that product distributions are an effective combining rule for Horn clause heads. We use a refinement of Pearl's loopy belief propagation [2] for the inference algorithm. We also extend the Kerst ing and De Raedt language by adding learnable distri butions. We propose a message passing algorithm based on Expectation Maximization [3] for estimating the learned parameters in the general case of models built in our system. We have also added some addi tional utilities to our logic language including second order unification and equality predicates.

Important Links:

Go Back