A General-Purpose Probabilistic Language

1.1 Introduction In a rational programming language, a program specifes a situation encountered by an agent; evaluating the program amounts to computing what a rational agent would believe or do in the situation. Rational programming combines the advantages of declarative representations with features of programming languages such as modularity, compositionality, and type systems. A system designer need not reinvent the algorithms for deciding what the system should do in each possible situation it encounters. It is sufficient to declaratively describe the situation, and leave the sophisticated inference algorithms to the implementors of the language. One can think of Prolog as a rational programming language, focused on computing the beliefs of an agent that uses logical deduction. In the past few years there has been a shift in AI towards specifications of rational behavior in terms of probability and decision theory. There is therefore a need for a natural, expressive , general-purpose and easy to program language for probabilistic modeling. This chapterpresents IBAL, a probabilistic rational programming language. IBAL, pronounced " eyeball " , stands for Integrated Bayesian Agent Language. As its name suggests, it integrates various aspects of probability-based rational behavior, including probabilistic reasoning, Bayesian parameter estimation and decision-theoretic utility maximization. This chapterwill focus on the probabilistic representation and reasoning capabilities of IBAL, and not discuss the learning and decision making aspects. High-level probabilistic languages have generally fallen into two categories. The first category is rule-based [Poole (1993); Ngo and Haddawy (1996); Kersting and de Raedt (2000)]. In this approach the general idea is to associate logic-programming-like rules with noise factors. A rule describes how one first-order term