3.3 Structure Learning

To learn the structure (clauses and weights) of an MLN, you use the learnstruct executable. The default structure learning algorithm is beam search, as in [3]. It can start from both empty and non-empty MLNs. When it starts from a non-empty MLN, it does not modify clauses that are derived from existentially quantified formulas or those containing variables with mutually exclusive and exhaustive values. Its options are similar to those of learnwts. In addition, it has options for controlling techniques that speed up the search. You can also restrict the types of clauses created during structure learning (see the developer's manual). To perform structure learning on the example data, one would use ALCHDIR/bin/learnstruct -i univ-empty.mln -o univ-empty-out.mln -t univ-train.db. Default values of various other parameters can be seen by running learnstruct without any arguments.