Output details
11 - Computer Science and Informatics
University of York
Online Bayesian inference for the parameters of PRISM programs
<24>The PRISM system is used to construct probabilistic computer programs. These probabilistic programs define arbitrarily complex probability distributions and so present a challenge to standard statistical methods for parameter estimation (whether Bayesian or non-Bayesian). This paper attacks the very difficult problem of online Bayesian analysis for PRISM programs. This requires maintaining a representation of an approximate posterior distribution which can be updated as new data arrives. The accuracy of the approximation is analysed using KL-divergence. The method is empirically tested using my own implementation of the PRISM system written in the Mercury language .