"Probabilistic Reasoning:Scaling Up"

1/30/1998: [time not recorded]
[location not recorded]

Abstract: Any application where an intelligent agent interacts with the real world must deal with the problem of uncertainty. Probability theory gives us a principled and coherent methodology for reasoning under uncertainty. In recent years, Bayesian networks have emerged as the dominant technology for making probabilistic reasoning a practical tool. Bayesian networks achieve effective knowledge representation and inference capabilities by utilizing a structural property that appears in many domains: the fact that, typically, each attribute is directly affected only by very few others. Bayesian networks have been used with great success in a wide variety of medium-scale applications. However, certain fundamental limitations on the expressive power of Bayesian networks renders them inadequate for handling large and complext domains. In the talk, I describe these limitations and discuss a new approach for scaling up probabilistic reasoning. Our new approach is based on an analogy between probabilistic modeling languages and programming languages. When viewed in this light, Bayesian networkds are analogous to logical circuits, clearly a far from optimal language for large scale applications. We show how various fundamental ideas from programming languages --- particularly function calls, encapsulation, and object-oriented programming --- can be imported into the framework of probabilistic modeling. The resulting language maintains the clean and coherent probabilistc semantics of Bayesian networks, while providing much greater expressivity. In particular, our framework allows us to deal with domains containing many objects; it also supports the representation of relations betweein objects, of hierarchically structured domains where objects can contain other objects, of classes of objects, and more. We also show that the additional structure encoded in these domain models can be exploited for inference. This property allows us to guarantee pe rformance scalability even for very large domain models. Finally, I discuss the connection between our representation language and more traditional knowledge representaiton languages, and show that our framework bridges the long-standing gap between the two main knowledge representation formalisms: Bayesian networks and frame-based logical representation. This talk covers joint work with Avi Pfeffer.

Last updated: Mon Jun 19 17:44:06 2006