An expert system is a “knowledge-based system that emulates expert thought to solve significant problems in a particular domain of expertise”.
What distinguishes expert systems from neural networks is that they are rule-based: the system holds a predefined body of knowledge used for every decision, and produces results by applying inference rules coded into it. Depending on its input and rules, an expert system can work as either a quantitative or a qualitative tool.
Knowledge base and inference engine
A typical expert system has two main modules: the knowledge base and the inference engine.
The knowledge base holds what the system knows about the domain it was designed for. A system working in finance, for example, would hold specific rules about decisions on shares. This knowledge is coded using a specific notation, usually in one of these forms:
- Rules
- Predicates
- Semantic nets
- Frames
- Objects
The inference engine processes and combines the facts of the problem at hand using the relevant part of the knowledge base, selected according to search criteria. How inference rules are written and applied varies greatly from system to system. Expert systems usually include other modules too, such as meta-knowledge.
The most important step is capturing the domain knowledge — the methods an expert actually uses to make the right decision.
That knowledge usually takes the form of heuristics, which are notoriously hard to put into words; the interviews needed to identify and collect them can take weeks.
Expert systems in practice
Little is published about these systems, since disclosing a successful approach could cost an operator its competitive advantage — and large sums of money. In general, operators now prefer neural networks for real-time forecasting, while expert systems are used more in areas of finance where the system must deliver a clear decision, such as validating credit-card transactions. They are also used in accounting, auditing and insurance decision-making.
