A Bayesian System for Integration of Algorithms for Real-time Bayesian Network Inference

Bayesian networks (BNs) are a key method for representation and reasoning under uncertainty in artificial intelligence. Both exact and approximate BN inference have been proven to be NP-hard. The problems of inference become even less tractable under real-time constraints. One solution to real-time AI problems is to develop anytime algorithms. Anytime algorithms are iterative refinement algorithms that trade performance for time. They improve the quality of the output as the amount of time increases. Another solution to real-time AI consists of metareasoning and integrating multiple approximation methods. To date, researchers have developed various exact and approximate BN inference algorithms. Each of these has different properties and works better for different classes of inference problems. Given a BN inference problem instance, it is usually hard but important to decide in advance which algorithm among a set of choices is the most appropriate. This problem is known as the algorithm selection problem. This dissertation proposal addresses the problem of real-time BN inference. It proposes work on both development of new anytime approximation algorithms and integration of multiple inference algorithms. Specifically, I first propose to study the multifractal properties of the underlying joint probability distributions (JPDs) of BNs and the skewness of the conditional probability tables (CPT) in real-world Bayesian networks. Based on these multifractal properties, I shall develop search-based anytime approximate inference algorithms for different characteristic classes of Bayesian networks. I then propose a Bayesian scheme for integration of various BN inference algorithms, including newly-designed search-based approximation algorithms. This scheme aims to integrate various BN inference algorithms into a unified framework so that the most appropriate inference algorithm would be selected for any given real-world inference problem specification. Novel features of the proposed research plan include: • The study of multifractal properties of the joint probability distributions represented by BNs • The design and implementation of search-based anytime approximate inference algorithms based on multifractal and skewness properties of characteristic BNs • The implementation of a random " real-world " Bayesian network generator to generate representative synthetic data for training and testing • An algorithm selection framework, itself a Bayesian learning and inference system that embodies expert knowledge about the domain of Bayesian network inference to select the best suitable algorithm for a given inference problem instance.