Grant: Discovery Grants Program - Individual

Bayesian networks are a means of exploiting conditional independence information in a joint probability distribution to facilitate both a more compact representation of the distribution, and faster inference from the distribution, without sacrificing the integrity of the original information. This is crucial for probabilistic decision-making in settings where variables interact in complex ways, such as medical diagnosis. It also facilitates and improves the accuracy of classification, forecasting, and fault diagnosis. Any software that improves the efficiency of decision-making, without sacrificing accuracy, makes deployment on consumer hardware more feasible. One of the major criticisms of using Bayesian networks in application is that their computational demands are often very high, both in terms of runtime and memory. Furthermore, many algorithms for computing over Bayesian networks use very sophisticated data structures and runtime libraries, which require further memory and sophisticated algorithms. Such implementations are not appropriate for simple architectures (e.g. portable devices), real-time systems (e.g. games) or memory-limited applications. In our proposed research program, we intend to continue development of streamlined data structures for computing probabilities from a Bayesian network at runtime. This approach will move much of the probabilistic computation offline, where the cost of computation is typically ignored, and minimizes costly runtime computation. The computation to be done at runtime will be packaged into a compiled version of the Bayesian network that makes reasonable demands on its host architecture, performing the required computation within the time and space constraints of the application.