A Method for Bayesian Meta-Reasoning Applied to Real-Time Systems Using Multiple Characterization

As Bayesian networks are applied to more complex and realistic real-world applications, the development of more efficient inference algorithms working under real-time constraints is increasingly important. In this paper, we present a method for meta-reasoning in Bayesian networks, which may be applied by real-time probabilistic systems that adopt anytime algorithms for approximate propagation of evidence and combine multiple simulation schemes. The proposed method is based on multiple characterizations of Bayesian networks to predict the algorithm that will provide the lower approximate error in future inferences, considering time restrictions. This method applies multiple regression analysis to create the conditional performance profiles of approximate inference algorithms. The analysis is based on experimental results to estimates of propositions that were used to create the utility curves of Gibbs Sampling and Stratified Simulation algorithms. These algorithms belong to stochastic and deterministic sampling methods, respectively. Some experimental analyses compare multiple and simple characterization of Bayesian networks for meta-reasoning and show better results in simulation errors when multiple characterizations are used.