First Generation Expert Systems
Many successful applications of AI were developed based on expert systems (ES) in the 70’s and 80’s. These AI systems were defined by Prof. Feigenbaum as “intelligent computer programs that use knowledge and inference procedures to solve problems that are difficult enough to require significant human expertise for their solution” [Feigenbaum82]. That is, an ES is a computer system that emulates the decision-making ability of a human expert [Giarratano98]. The terms expert systems, knowledge-based system (KBS), and rule-based system (RBS) are often used synonymously. These systems are usually composed of a knowledge base (rules and facts), an inference engine, and a user interface. As examples of successful and important ES’s, we can cite: MYCIN (Shortliffe, et. al.), DENDRAL (Feigenbaum), PROSPECTOR (Duda, Hart, et. al.), and XCON/R1 (McDermott). Several ES shells were developed in order to simplify and accelerate the development of these systems like CLIPS (Giarratano), JESS (Sandia Labs.), Eclipse (Haley Enterprise), ExSys (Multilogic Inc.), OPS5, and some PROLOG/LISP tools.
Expert Systems Drawbacks – The First Generation
The first generation ES’s presented some problems related mainly to knowledge acquisition and reasoning under uncertainty. In fact, the problem of transferring human knowledge into an ES is so major that it is called “knowledge acquisition bottleneck” [Giarratano98]. This is a descriptive term because the knowledge acquisition bottleneck constricts the building of an ES like an ordinary bottleneck constricts fluid flow into a bottle. It’s a hard work (long and difficult) for an expert to explain (all) his knowledge and reasoning used to solve problems in a specific domain, and then code all this knowledge into facts and rules. Several techniques were developed in order to achieve this task, but no one can really explain all the knowledge used by the expert to solve a problem. Human reasoning is a complex task and we need to take into consideration that all knowledge – even though related to a very small problem - can’t be obtained and/or may not be available from the expert (due to common sense, particular cases, implicit context relations, etc).
The second main problem of ES’s is how to consider inexact and approximate knowledge, as usually, these systems use only ‘binary’ inputs and ‘binary’ decisions. For example, can it infer that someone is sick if his body temperature is around 39oC, and conclude that he is a quite sick or very sick? This problem is related to the fact that human knowledge is mainly incomplete, inexact, and approximate.
When you try to put together knowledge bases obtained from different sources, almost always you will get some contradictory and incompatible rules. Even if you only got the knowledge from one expert, this problem normally appears.
First Generation Expert Systems
From Expert Systems to Hybrid Systems – The Second Generation
The second generation ES’s are characterized by the addition of automatic knowledge acquisition tools (machine learning tools) in order to minimize the “knowledge acquisition bottleneck”, and of additional features and techniques to allow these systems to reason under uncertainty.
Machine learning tools [Mitchell97, Nilsson98] are used in ES’s to automatically acquire knowledge from sources other than the expert. These tools allow us to improve the system knowledge base, learning from available data about a specific problem. For example, in a medical diagnosis application, we can acquire knowledge from hospitals’ patient records. Several machine learning algorithms were proposed in order to improve knowledge acquisition from cases (examples): induction of decision trees (IDT) [Quinlan93], artificial neural networks (ANN) [Arbib95, Rumelhart86], GA [DeJong88, Golberg89, Davis91] and CBR [Kolodner93]. Each of these tools has advantages, and also some drawbacks.
Reasoning under uncertainty is very important in dealing with the actual world, practical applications, and expert knowledge. In order to consider uncertainty and also increase the knowledge representation of ES’s, extensions to traditional symbolic systems were proposed. These extended ES’s attempt to incorporate some concepts from the theory of classical probability to reason under uncertainty, but avoiding some imposed restriction like the “closed-world assumption” [Giarratano98].
Classical propositional and rule based systems were extended adopting certainty factors (originally developed for the MYCIN system), theory of evidence (Dempster-Shafer theory), approximate reasoning (FL), theory of possibility, Bayesian inference, Markov chains, and fuzzy petri-nets. Although these approaches address some problems related to representation and manipulation of uncertain, inexact, and approximate knowledge, they have difficulties of specifying the amount of uncertainty, confidence, possibility, probability, belief, and disbelief on facts and rules.
Second Generation Expert Systems
The second generation ES’s attempt to put together tools for:
Connectionist Symbol
Processing:
CLARION and CONSYDERR [Sun], RAAM [Pollack, Sperduti], DCPS [Touretzky and Hinton],
CSN [Shastri]
Neuro-GA
Neuro-Fuzzy: Fynesse, ANFIS, NEFCON, NEFCLASS, FAGNIS, FINEST
Neuro-IDT
Neuro-CBR: Probis [Malek]
Neuro-KBS: CORE [Kasabov], Synhesys [Giacometi], KBCNN [Fu], KBANN [Towell], INSS [Osorio], MACIE[Gallant], EXPSYS [Sima], Rapture[Mahoney]
Neuro-DFA:
Others:
EBNN [Mitchell], SCANDAL [Melanie], Honavar, Obradovic, Medsker, Pomerleau
