Marek J. Druzdzel, Air Force Office of Scientific Research (AFOSR), "Canonical Probability Distributions for Model Building, Learning, and Inference", 2003--2005 (three years), total amount $381,883

The project is a continuation of a New World Vistas projects funded by AFOSR that resulted in theoretical contributions as well as the development of a general purpose environment for decision-theoretic modeling, SMILE and its Windows user interface, GeNIe (available for download at http://www.sis.pitt.edu/~genie).  Development of the software, used in a number of academic and industrial centers across the world, has contributed to a popularity of the decision-theoretic methods in intelligent systems. At the moment, a major obstacle to wide dissemination of these methods is the considerable effort that goes into constructing probabilistic models.  Full specification of a general discrete conditional probability distribution of a node N in a network requires a multitude of numerical parameters that grows exponentially with the number of parents of N.  One effective way of dealing with this problem is through use of parametric distributions that are capable of approximating conditional probability distributions with fewer parameters.  And so, existing Noisy-OR or Noisy-MAX parametric distributions, while offering very good approximations, reduce the number of parameters from exponential to linear in the number of parents.  We focus on (1) studying and refining existing parametric probability distributions, such as Noisy-OR and Noisy-MAX distributions, (2) developing other intuitive parametric probability distributions, (3) studying how the parametric distributions can be used in learning discrete conditional probability distributions from data, (4) studying how belief updating algorithms for Bayesian networks can take advantage of parametric probability distributions, (5) continuing our previous work on efficient stochastic sampling algorithms, (6) developing special features that will support diagnostic applications, and (7) demonstrating the power of the modeling environment in practical applications involving diagnosis of complex systems, such as medical diagnosis and diagnosis of engineering devices.  Our theoretical contributions will be implemented in SMILE and GeNIe.