Multivariate Analysis Applied in Bayesian Metareasoning

As Bayesian networks are applied to more complex and realistic real-world applications, the development of more efficient inference approaches is increasingly important. This paper presents a method for metareasoning in Bayesian networks adopting prediction models to select algorithms for the inference tasks, when multiple schemes are used to calculate the propagation of evidence. The proposed method is based on multiple characterizations of Bayesian networks and prediction models to select the algorithm that will provide the best performance in future inferences. Logistic regression analysis is applied to determine when exact algorithms may be used for specific tasks. The prediction models of approximate inference algorithms are created by multiple regression analysis, based on experimental results using Variable Elimination, Gibbs Sampling and Stratified Simulation algorithms. These algorithms belong to exact method, stochastic and deterministic sampling methods, respectively. Experimental analyses compare some alternative models and show better results when multivariate analysis is applied. Key-Words: Bayesian, Metareasoning, Logistic Regression, Multiple Regression, Inferences, Prediction Models