Knowledge Representation for Expert Systems

The purpose of this article is to summarize the state-of-the-art of the expert systems research field. First, we introduce the basic notion of knowledge, and specifically of shallow knowledge, and deep knowledge. The first section of the document summarizes the history of the field. We analyze the differences between the first generation of expert systems, based primarily upon rulebased and frame-based representation of shallow knowledge. We mainly concentrate on the most important expert systems and their impact on the subsequent research. These are the traditional Mycin and Prospector expert systems, but also less famous ones such as General Problem Solver, Logic Theory Machine, and others. Finally, we present some modern expert systems and shells such as Gensym’s G2 and also some light-weight prolog based expert systems, usually based on deep knowledge of the domain. In the forth section, we compare various knowledge representation languages. We briefly describe each, present some inference techniques, and also discuss primary the upsides and downsides. For each, we finally present successful expert systems and shells using the language. As for shallow knowledge, we review mainly rule-based and framebased knowledge representation languages. We try to argue why these are not very suitable to model complex relationships present in many real world applications and therefore not suitable for deep knowledge representation. Subsequently, we present early semantic networks as the first attempt to model deep knowledge. Then more in depth, we analyze the approaches based on simplifications and extensions of traditional logic. In the first place this is the propositional logic, first order predicate logic, modal logic, and finally logic programming (Prolog). We further continue extended with constraint programming. Then we discuss nonmonotonic knowledge representation languages, such as answer set programming and default logic. At last we analyze representation of knowledge for continuous domains, mostly addressed by qualitative and semi-qualitative simulation. The fifth section presents various uncertainty measures and their combinations with the previously mentioned representation languages. First we reason why we need to represent uncertainty. Then we present generally required properties to which the measure should adhere. We follow with the classic probability theory and its combinations with propositional, first order predicate logic, and modal logic, and logic programs. We discuss the most popular representation model, Bayesian belief networks. We point out the properties and reasons why probability is a measure that works perfectly for statisticians, but is not completely satisfactory for many artificial intelligence domains. We continue with Dempster-Shafer theory. We introduce the Transferable Belief Model that employs this measure. Next we present possibility theory and it’s combination with predicate logic, known as fuzzy logic. We present why this seems to be a very popular choice for simple systems and why it seems unsuitable for large and complex expert systems. We also presents systems trying to combine rule-based systems with neural networks.