Probabilistic Logic Programming can be used to model domains with complex and uncertain relationships among entities. While the problem of learning the\udparameters of such programs has been considered by various authors, the problem\udof learning the structure is yet to be explored in depth. In this work we present an\udapproximate search method based on a one-player game approach, called LEMUR. It\udsees the problem of learning the structure of a probabilistic logic program as a multiarmed bandit problem, relying on the Monte-Carlo tree search UCT algorithm that\udcombines the precision of tree search with the generality of random sampling. LEMUR\udworks by modifying the UCT algorithm in a fashion similar to FUSE, that considers a\udfinite unknown horizon and deals with the problem of having a huge branching factor.\udThe proposed system has been tested on various real-world datasets and has shown\udgood performance with respect to other state of the art statistical relational learning\udapproaches in terms of classification abilities.
展开▼
