This Small Business Innovation Research Phase I project is to develop a new framework for writing and executing stochastic programs. The work expands and generalizes the stochastic functional language originally proposed by Koller, McAllister, and Pfeffer at Stanford University. The goal is to implement a functional object oriented stochastic language for expressing probabilistic relationships over variables and objects. The output includes both joint and conditional probability distributions. The language provides a compact representation of Bayesian Belief Networks as well as a larger class of stochastic models. It enhances the power of Bayesian Belief Networks in several respects. It is more expressive, providing mechanisms for representing abstract relationships as well as functions for expressing model composition and model transformation. The language also supports an inference algorithm that is more efficient than standard Bayesian Belief Network algorithms. It achieves efficiency through lazy evaluation, context-sensitive inference, and compact factorization of conditional probability tables. This new language is an evolutionary step forward from Bayesian Belief Network formalisms in both expressive power and inference and will support stochastic modeling in entirely new domains of science and engineering. Stochastic modeling and stochastic inference are useful in design, simulation and diagnosis, as well as in multiple aspects of decision support. If successful, this new generation of stochastic programming tools will have potential applications in: advanced engineering design; clinical data interpretation and diagnostic systems in medicine; diagnosis and predictive monitoring in industrial control systems; Market trend analysis and prediction for financial institutions; decision support for business, government and other organizations; scientific modeling of stochastic systems, and; genealogical analysis.
