Uncertainty Handling in Expert Systems: From Certainty Factors to Fuzzy Logic

Real-world knowledge is uncertain. Symptoms suggest diagnoses without confirming them. Sensor readings contain noise. Expert opinions conflict. Any expert system operating on real data must handle uncertainty explicitly — and the choice of uncertainty formalism significantly affects system behavior and maintainability.

Why Certainty Factors Are Still Used

MYCIN's certainty factors (CF) — a confidence score from -1 to +1 associated with each fact and rule — remain common despite theoretical criticism. The reason is pragmatic: domain experts can assign CF values intuitively, and the combination rules (CF(A AND B) = min(CF(A), CF(B))) are simple enough to reason about manually.

CF is appropriate when:

• Precise probability estimates are unavailable
• Domain experts think in terms of "strong evidence", "weak evidence", "contradicting evidence"
• The system needs to explain its confidence level in plain terms

CF is inappropriate when:

• Precise probability estimates are available from data
• Independence assumptions implicit in CF formulas are violated (common in practice)

Bayesian Networks

A Bayesian network represents variables and their conditional dependencies as a directed acyclic graph. Given observed evidence, Bayesian inference computes posterior probabilities for unobserved variables.

Advantages: mathematically rigorous, handles complex dependencies, updateable as new evidence arrives.

Disadvantages: requires quantified probability estimates (hard to elicit); structure learning from data requires large datasets; exact inference is NP-hard in general graphs.

Use Bayesian networks when you have calibrated probability data and the domain structure is well-understood.

Dempster-Shafer Theory

Dempster-Shafer (DS) theory generalizes Bayesian probability by allowing explicit representation of ignorance — distinguishing "I don't know" from "50-50". Evidence supports not just individual hypotheses but sets of hypotheses.

DS is particularly useful in diagnostic systems where the set of possible diagnoses may not be exhaustive and some evidence patterns do not point to any single diagnosis.

The major limitation is computational: combining many pieces of evidence becomes expensive, and the DS combination rule can produce counterintuitive results when evidence sources conflict strongly.

Fuzzy Logic

Fuzzy logic handles vagueness in concept membership rather than uncertainty about facts. In crisp logic, a patient is either "elderly" (≥65) or not. In fuzzy logic, a 64-year-old has 0.9 membership in "elderly" and a 50-year-old has 0.3 membership.

Fuzzy rules fire to a degree proportional to input membership:

IF temperature IS high AND pressure IS moderate THEN risk IS elevated

Fuzzy systems excel at:

• Control systems where linguistic rules from operators are natural (automotive, HVAC)
• Systems with continuously-varying inputs where crisp thresholds would produce unstable behavior

Fuzzy logic does not handle epistemic uncertainty (unknown facts) — it handles linguistic imprecision.

Choosing an Approach

Conclusion

No single uncertainty formalism is universally superior. The right choice depends on what kind of uncertainty dominates your domain, what data is available, and how experts naturally express their knowledge. Mixing formalisms in a single system is possible but introduces complexity; prefer one primary approach with targeted use of others where clearly justified.

Keywords: certainty factors, Bayesian networks, Dempster-Shafer, fuzzy logic, uncertainty handling, expert systems, knowledge-based systems, probabilistic reasoning, MYCIN certainty factors, imprecise knowledge