| A key issue throughout the history of Artificial Intelligence (AI) is the problem of knowledge representation. During the early days of AI, logic provided the basis for most knowledge engineering. The logic-based approaches were found to be quite brittle and couldn't handle effectively the uncertainty that exists everywhere in the world.; In the late 1980s and early 1990s, Bayesian Networks (BNs), invented by Judea Pearl, came into vogue. By providing a principled probabilistic foundation for handling uncertainty, these Networks overcame many of the weaknesses of the earlier knowledge representations. However, BNs themselves are not as expressive as is needed to handle the true complexity of the world. BNs are propositional in nature; they can only express models containing a finite number of variables. This limitation makes it cumbersome to express relations across classes of variables.; Thus, there is a need for first-order representations of stochastic models. These first-order systems are aimed at problems with repetitive or recursive structure. This thesis explores such first-order modeling systems from the perspective of three important programming language paradigms.; The first paradigm is object-oriented programming. The stochastic equivalent uses objects to permit encapsulation as well as infinite recursion. The second paradigm is functional programming. Here the functions from λ-calculus can be extended to allow arbitrarily complex stochastic mappings. The final paradigm is based on logic programming. This is the most direct lifting of propositional BNs to a first-order predicate representation.; After presenting these three first-order languages, this thesis compares these approaches and discusses their relative utility for first-order stochastic modeling. |
