Expert Systems Questions and Answers

Answer: C

Conflict is incompatible conclusions from the same conditions, redundancy is duplicated effect, and unreachability is a rule no combination of facts can trigger. Circularity and subsumption complete the standard list. Static verification tools search for all five, which is the practical response to the discovery that rule bases do not remain modular as they grow.

Answer: D

Because the knowledge could not be specified in advance, a small prototype was built, shown to the expert, and refined through many cycles. Seeing the system get a case wrong was often the only way to surface a missing rule. Waterfall and formal specification presuppose a requirements document that does not exist here, and modular parallel work founders on the interactions between rules.

Answer: A

Knowledge ages as the domain changes, and a rule base with thousands of interacting rules is expensive to keep correct, which is what XCON's permanent maintenance team demonstrated. Hardware improved rapidly, shells were plentiful by the mid eighties, and experts generally cooperated. Maintenance cost, rather than any failure of capability, is the standard explanation for the retreat that followed. Checking the units of each quantity catches a wrong formula before the arithmetic begins.

Answer: B

Inducing the knowledge from examples sidesteps elicitation entirely, which is why learned models displaced hand-built rule bases in most diagnostic applications. Faster engines, better editors and interchange formats each helped at the margin without touching the underlying difficulty. The trade is interpretability: a learned model rarely explains itself as a rule trace does, which is why explanation is now a research topic again. Checking the units of each quantity catches a wrong formula before the arithmetic begins.

Answer: C

A fired rule chain reads as an argument, whereas the weights of a learned model do not, which is the trade accepted when induction replaced elicitation. Learned models handle partial input, evaluate quickly and improve with data, so those are not the losses. The current interest in explainable models is an attempt to recover what the rule-based era had by construction.

Answer: D

Tacit knowledge is precisely what an expert cannot articulate on request, and it is the reason interviews alone are insufficient. Protocol analysis and observation of real cases are the standard responses. Meta-knowledge concerns strategy, shallow knowledge concerns depth, and compiled knowledge is a related but narrower term. Naming the obstacle correctly is the first step in choosing an elicitation method. Working one numerical case through both cases side by side fixes the distinction better than any rule of thumb.

Answer: A

Meta-knowledge governs which rules to try, when to abandon a line and how to organise the search, and making it explicit lets the system explain and teach its strategy. TEIRESIAS and NEOMYCIN both did so. Tacit, deep and procedural knowledge name other distinctions. The move from implicit control to declared strategy is one of the field's clearer intellectual advances.

Answer: B

A system gives the same answer to the same case at any hour, never tires and can be copied, which is the practical argument for building one. Recognising its own limits, applying common sense and learning from experience are exactly the things it does not do, and the last of these is what case-based reasoning and machine learning were introduced to supply.

Answer: C

A human recognises unfamiliar ground and reasons cautiously from first principles, whereas a rule base answers confidently or not at all. Consistency, tirelessness and speed favour the machine, and rule traces actually give better step-by-step justification than most experts produce. The asymmetry at the boundary of competence is the one that has proved hardest to eliminate.

Answer: D

DENDRAL's plan step derived constraints from the spectrum and fed them into the structure generator, so impossible candidates were never enumerated at all. Generating everything and testing afterwards would have been hopeless for molecules of any size. Ranking, parsing and human review happen at other points and do not perform the pruning. The lesson that constraints belong inside the generator rather than after it recurs throughout artificial intelligence, from constraint satisfaction to modern program synthesis.