Informatik, TU Wien

What Is a Rule in Artificial Intelligence?

If-then rules, which are arguably the most common form of knowledge representation in Artificial Intelligence, are ambiguous.


If-then rules, which are arguably the most common form of knowledge representation in Artificial Intelligence, are ambiguous. They can be interpreted both as logic programs having the form if conditions then conclusions and as production rules having the form if conditions then do actions. The relationship between these different kinds of rules has received little attention in the AI literature; and, when it has, different authors have reached entirely different conclusions.

Some authors, such as Russell and Norvig in their textbook Introduction to Artificial Intelligence, view production rules as just logical implications used to reason forward, while Herbert Simon in the MIT Encyclopedia of Cognitive Science views the logic programming language Prolog as one of many production system languages. On the other hand, Thagard in his Introduction to Cognitive Science denies any relationship between logic and production rules at all.

In this talk, I will explore the relationships between logic programs and production rules and propose a framework that combines the two kinds of rules and eliminates the overlap between them. The framework uses production rules for sentences of the form if conditions then achieve goals, and it uses logic programs both to evaluate conditions and to achieve goals by reducing goals to sub-goals. I will discuss the problems of giving the resulting framework both a declarative, model-theoretic semantics and an operational semantics in the form of a transition system.


Robert Kowalski studied at the University of Chicago, the University of Bridgeport, Stanford University, and the University of Warsaw, before completing his PhD at the University of Edinburgh in 1970. He was a Research Fellow at the University of Edinburgh from 1970 to 1975, and has been at Imperial College London since 1975. He is a Fellow of the Association for the Advancement of Artificial Intelligence, the European Co-ordinating Committee for Artificial Intelligence, and the Association for Computing Machinery.

Kowalski began his research in the field of automated theorem-proving, developing both SL-resolution with Donald Kuehner and the connection graph proof procedure. He is best known for his contributions to the development of logic programming, starting with the procedural interpretation of Horn clauses. He developed the minimal model and the fixpoint semantics of Horn clauses with Maarten van Emden, and the event calculus and the application of logic programming to legal reasoning with Marek Sergot.

Kowalski was one of the developers of Abductive Logic Programming. This led to his collaboration with Phan Minh Dung and Francesca Toni, showing that most logics for default reasoning can be regarded as special cases of assumption-based argumentation. It also led to the development with Fariba Sadri of the ALP agent model, in which beliefs are represented by logic programs and goals by integrity constraints. His current research focuses on the application of Computational Logic to Cognitive Science. A draft of a recent book "How to be Artificially Intelligent" can be found at