4.25 Expert systems: architecture, knowledge engineering and rule-based inference

Work through these

• Say what an expert system is, and how it differs from a conventional program

  The knowledge sits in data structures rather than in the code, so a different problem in the same domain can reuse the program. A conventional program has to be rebuilt when what it knows changes.

• List the phases of building one: identification, conceptualisation, formalisation, implementation, testing

  Five stages, from deciding the scope with a domain expert to evaluating a prototype and revising it. Naming them matters because most failures happen in the first two and get blamed on the last.

• Describe the architecture: knowledge base, inference engine, working memory, user interface, explanation module

  Static rules and facts on one side, the consultation's own growing facts on the other, and an engine that searches both. The explanation module is what lets a user ask how and why.

• Write production rules, and run forward chaining over them

  If-then rules applied from the known facts towards whatever they imply, until a goal appears. It suits situations where the data arrives first and the question comes later.

• Run backward chaining, and say which problems suit which direction

  Starting from a hypothesis and looking for the facts that would support it. Diagnosis is the natural fit, which is why the medical systems of the period worked this way round.

• Describe knowledge acquisition, and why it is the hard part

  Getting what an expert knows into rules, through interviews, texts and case histories, and keeping it current afterwards. This is where most projects of the period ran aground.

• State the limits: no common sense, no creative response, no adaptation without new rules

  The system cannot handle a situation outside its rules, cannot invent, and changes only when someone edits the knowledge base. Those limits are the reason the second winter arrived.

• Judge whether a problem suits an expert system at all

  Specialised knowledge, a high payoff, willing experts, and a cost the result can justify. Asking this first is the check the field skipped when the approach was fashionable.