Learning failure-free PRISM programs

• Main Author: Alsanie, Waleed
• Other Authors: Cussens, James
• Published: University of York 2012
• Subjects: 005
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.568108

First-order logic can be used to represent relations amongst objects. Probabilistic graphical models encode uncertainty over propositional data. Following the demand of combining the advantages of both representations, probabilistic logic programs provide the ability to encode uncertainty over relational data. PRISM is a probabilistic logic programming formalism based on the distribution semantics. PRISM allows learning the parameters when the programs are known. This thesis proposes algorithms to learn failure-free PRISM programs. It combines ideas from both areas of inductive logic programming and learning Bayesian networks. The learned PRISM programs generalise dynamic Bayesian networks by defining a halting distribution over the sampling process. Each dynamic Bayesian network models either an infinite sequential generative process or a sequential generative process of a fixed length. In both cases, only a fixed length of sequences can be sampled. On the other hand, the PRISM programs considered in this thesis represent self-terminating functions from which sequences of different lengths can be obtained. The effectiveness of the proposed algorithms on learning five programs is shown.