Expert Systems and Uncertainty Interview Questions: Model Answers with Worked Reasoning

Prepare for expert-system follow-ups with a deterministic rule trace, two Bayesian updates, a certainty-factor calculation and an auditable design answer.

KnowledgeGate Team

Exam prep & CS education

Updated 18 Nov 20266 min read
Expert Systems and Uncertainty Interview Questions: Model Answers with Worked Reasoning

Memorising that an expert system contains a knowledge base and an inference engine may secure a definition, but it does not prepare you for follow-up questions on rule firing, explanations, conflicting evidence or uncertain conclusions. An auditable vehicle-diagnosis rule trace updates a Bayesian belief from 10% to 30.8% and then 57.1%, and combines MYCIN-style certainty factors to reach 0.692. For broader interview context, Resume & Interview Preparation places this technical discussion within a wider preparation route.

Expert systems interview question: what is an expert system?

Model answer: An expert system is a knowledge-based AI program that represents domain expertise explicitly, applies an inference procedure to case facts, produces a recommendation or diagnosis, and can expose its reasoning path.

Its five working parts become clear in a vehicle case. The knowledge base stores rules. Working memory, or the fact base, begins with no_start=true, slow_crank=true, dim_lights=true and fuel_available=true. The inference engine matches those facts to rules. The explanation facility records what fired and why. The knowledge-acquisition interface lets an engineer add or revise rules.

Aspect

Expert system

Conventional program

Machine-learning model

Where knowledge lives

Explicit rules

Procedural code

Learned parameters

How behaviour changes

Edit rules

Edit code

Retrain or update

Explanation

Rule trace

Code path

Model-dependent explanation

Each approach fits different constraints; none is always superior.

Knowledge representation and inference: trace rules before naming the algorithm

Use R1: IF slow_crank AND dim_lights THEN weak_battery, R2: IF weak_battery THEN inspect_battery_terminals, and R3: IF no_start AND fuel_available AND NOT weak_battery THEN inspect_ignition. From the four initial facts, R1 adds weak_battery; R2 then adds inspect_battery_terminals. R3 does not fire because NOT weak_battery is false.

Forward chaining is data-driven: it starts from facts and suits monitoring or cases where several conclusions may help. Backward chaining is goal-driven: starting with weak_battery, it asks whether slow_crank and dim_lights can be established. If several rules are eligible, specificity, priority or recency may resolve the conflict, but the system must define its policy. Knowledge Representation in AI: Logic, Frames and Rules develops the representation choices behind such rules.

The trace shows how updated working memory enables one rule and blocks another before any algorithm label is applied. The engine continues matching until no eligible rule remains.

Diagram of a vehicle expert-system trace: initial facts trigger R1 and R2, while R3 is blocked because weak_battery is true.

Uncertainty in expert systems: probability, certainty factors and fuzzy truth

Uncertainty enters through noisy sensors, incomplete reports, heuristic rules or conflicting observations. It concerns whether a proposition is true; fuzziness concerns how strongly a graded term such as temperature is high applies.

Method

What it represents

Good fit

Limitation

Bayesian probability

Calibrated belief under stated conditional probabilities

Credible priors and likelihoods

Needs defensible estimates

Certainty factors

Expert confidence under a declared scheme

Transparent heuristic rules

Not calibrated probability

Fuzzy logic

Graded membership, such as high_temperature = 0.7

Vague boundaries

Membership 0.7 is not automatically 70% probability

Dempster-Shafer theory

Support for sets of hypotheses, with mass left for ignorance

Evidence with uncommitted belief

Combination can be difficult under conflict

Choose Bayes for credible probabilities, certainty factors for heuristic confidence, fuzzy logic for vague terms, and Dempster-Shafer when set-valued support matters. AI & ML for Placements offers a broader structured route through AI concepts.

Bayesian uncertainty interview question: update a weak-battery diagnosis

State the constructed assumptions first. Let B mean weak battery, with P(B)=0.10 and P(not B)=0.90. Let D mean dim headlights, with P(D|B)=0.80 and P(D|not B)=0.20.

Bayes gives:

P(B|D) = [0.80 x 0.10] / [(0.80 x 0.10) + (0.20 x 0.90)]

= 0.08 / (0.08 + 0.18) = 0.08 / 0.26 = 0.3077, or 30.8% to one decimal place.

Now add S = slow crank. For this toy calculation only, assume D and S are conditionally independent given B and given not B, with P(S|B)=0.90 and P(S|not B)=0.30.

P(B|D,S) = [0.10 x 0.80 x 0.90] / [(0.10 x 0.80 x 0.90) + (0.90 x 0.20 x 0.30)]

= 0.072 / (0.072 + 0.054) = 0.072 / 0.126 = 0.5714, or 57.1%.

The denominator normalises the weak-battery and not-weak-battery evidence paths, so the posterior remains coherent. Evidence changes belief through a likelihood ratio, not by adding percentages. The second result depends on conditional independence. If that assumption fails, use joint likelihoods P(D,S|B) and P(D,S|not B). This is a ranked belief, not a guaranteed diagnosis.

Two-panel Bayesian update diagram showing dim-headlight evidence raising weak-battery belief to 30.8%, then slow-crank evidence to 57.1%.

Certainty-factor interview question: combine two heuristic rules

Use a classic MYCIN-style toy convention, not probability. Facts have CF(slow_crank)=0.90, CF(dim_lights)=0.80 and CF(voltage_below_11.8V)=0.60.

  • R1: IF slow_crank AND dim_lights THEN weak_battery, rule CF =0.70

  • R2: IF voltage_below_11.8V THEN weak_battery, rule CF =0.50

For R1, take the minimum premise confidence: min(0.90,0.80)=0.80. Its conclusion contribution is 0.80 x 0.70=0.56. R2 contributes 0.60 x 0.50=0.30.

The minimum makes the weaker conjunct limit R1's premise support. Multiplication then discounts that support by the rule's own CF. The positive combination adds R2 only to the support space left after R1.

For two positive contributions, declare CFcombined = CF1 + CF2(1-CF1):

0.56 + 0.30(1-0.56) = 0.56 + 0.30(0.44) = 0.56 + 0.132 = 0.692.

The toy action threshold is 0.65, so 0.692 >= 0.65 triggers recommend battery inspection. A useful explanation reports both rules, premise confidences, contributions 0.56 and 0.30, the combination formula, final 0.692 and threshold 0.65. The result is not calibrated 69.2% probability, and this threshold is not universal. Negative evidence needs the system's separately declared conflict formula.

Expert-system design question: make the answer maintainable and auditable

Interview prompt: Design a support expert system for this vehicle case with 500 rules, 40 observable facts and three possible recommendations.

A strong design defines a rule schema, typed fact validation, an agenda of eligible rules, deterministic conflict resolution, explanation traces, rule versioning, and tests for expected conclusions and rules that must not fire. Partition the 500 rules by subsystem and index premises so a new fact does not scan every rule. Store each inferred fact's provenance as (rule_id, input_fact_ids, rule_version).

Use a confidence threshold only after documenting its semantics and calibration. Route low-confidence cases or incompatible conclusions to human review. The knowledge-acquisition bottleneck is often harder than inference: teams must elicit tacit expertise, resolve disagreement and keep rules current. Assign review ownership, test every rule change, and monitor for stale or contradictory rules.

How service and product interviews probe expert systems

These are useful forms, not employer-process promises. Fundamentals rounds can test definition, components and chaining. Deeper rounds can require traces, model choices, defended assumptions, user explanations, maintenance or human escalation.

Weak answer

Why it fails

Stronger repair

Expert systems learn rules automatically

Rules are explicit

A separate learning component may update them

Fuzzy value 0.7 means 70% likely

Membership differs from probability

Define membership

Bayes adds evidence percentages

Bayes normalises prior-weighted likelihoods

Show numerator and normaliser

CF 0.692 means 69.2%

CF depends on its declared model

State the combination convention

A fired rule proves the diagnosis

Inference depends on facts and rules

Expose the trace and assumptions

For retrieval practice, try the companion Expert Systems and Uncertainty MCQs: 11 Solved Questions. The set does not show employer frequency.

Expert systems and uncertainty: the short version and next step

  • Knowledge is explicit.

  • Inference applies rules to case facts.

  • Forward chaining starts from data.

  • Backward chaining starts from a goal.

  • Uncertainty needs declared semantics and combination rules.

Recheck R1 -> weak_battery -> R2; Bayes 10% -> 30.8% -> 57.1% under the stated assumptions; and 0.56 plus 0.30 combining to 0.692, above the toy threshold 0.65. Ten-minute drill: give the 45-second definition, redraw the trace, recompute both examples, then distinguish CF from probability. The Interview & Resume Preparation Course provides a structured route.