Flexible causal discovery with MML

Causal Models, in the form of Bayesian Networks, are an increasingly important tool for modeling and reasoning under uncertainty in numerous domains. Their ability to represent and reason about causal relationships in a humanly intuitive fashion explains this growing importance. Traditionally Bayesian Networks have been manually constructed, though over the last decade there has been a growing interest in automated discovery. In this dissertation, two major topics are examined while focusing on improving automated discovery of causal models. Firstly a new measure, Causal KL (CKL), is developed for evaluating causal discovery algorithms. CKL is more appropriate than regular KL for assessing causal discovery as it is sensitive to the “causal” part of a Bayesian network, while KL is unable to distinguish between observationally equivalent models. Secondly, two new enhancements to the CaMML (Causal MML) program are developed; (1) to allow flexible learning of local structure, and (2) learning from a combination of raw data and expert elicited information. Both additions provide large improvements to the quality of models produced by CaMML.