Toward Markov Logic with Conditional Probabilities.
Jens Fisseler · 2008
Combining probability and first-order logic has been the sub-ject of intensive research during the last ten years. The most well-known formalisms combining probability and some sub-set of first-order logic are probabilistic relational models (PRMs), Bayesian logic programs (BLPs) and Markov logic networks (MLNs). Of these three formalisms, MLN is the currently most actively researched. While the subset of first-order logic used by Markov logic networks is more expressive than that of the other two formalisms, its probabilistic seman-tics is given by weights assigned to formulas, which limits the comprehensibility of MLNs. Based on a knowledge rep-resentation formalism developed for propositional probabilis-tic models, we propose an alternative way to specify Markov logic networks, which allows the specification of probabili-ties for the formulas of a MLN. This results in better com-prehensibility, and might open the way for using background knowledge when learning MLNs or even for the use of MLNs for probabilistic expert systems.