Expert system development steps

An expert system is a program that captures the knowledge of human experts in a specific domain and uses it to give expert-level advice or solutions (e.g. MYCIN for diagnosing blood infections).

Stages of development

 1 Problem identification
        v
 2 Feasibility study / requirements
        v
 3 Knowledge acquisition  <------------+
        v                              |
 4 Knowledge representation            | refine
        v                              |
 5 Select tool / design and build -> 6 Testing and validation
        v
 7 Implementation / deployment
        v
 8 Maintenance

1. Problem identification: decide whether an expert system suits the problem (needs a narrow domain, a scarce expert, rules and heuristics, and a clear benefit).
2. Feasibility and requirements: study cost, benefit, available experts, hardware and users; set goals.
3. Knowledge acquisition: the knowledge engineer collects knowledge from experts by interviews, observation and documents.
4. Knowledge representation: encode the knowledge as rules, frames or semantic nets, and choose a reasoning strategy (forward or backward chaining).
5. Design and development: choose a shell (e.g. CLIPS, Prolog) or language; build the knowledge base, inference engine, interface and explanation facility; build a prototype.
6. Testing and validation: run test cases and compare with the expert's decisions; correct rules; repeat steps 3-6 (incremental refinement).
7. Implementation: install, train users and integrate with existing systems.
8. Maintenance: update the rules as the domain changes.

Example: car fault diagnosis

• Problem: car does not start; experts (mechanics) are few.
• Knowledge acquisition: interview a mechanic and obtain rules.
• Representation: IF battery_voltage is low THEN fault = flat_battery. IF engine_cranks AND no_spark THEN fault = ignition.
• Build with a shell; the system asks the user questions (does the engine crank?).
• Test with 50 past cases; fix wrong rules; deploy in the workshop; add new models later.