Certainty Factors and Rule-Based Uncertainty (Overview) — Beginner-Friendly Notes

1. Why This Topic Exists (historical patch)

Classic rule-based systems (MYCIN) were logical but brittle — they needed to handle "maybe". Before full Bayes became practical, they invented Certainty Factors (CFs) as a heuristic blend of belief and disbelief.

Think: CF = quick, modular uncertainty for if-then rules before you learn Bayes nets.

2. Definition (one line)

For hypothesis $H$ given evidence $E$:

• Measure of Belief $MB$: increase in belief in $H$ due to $E$.
• Measure of Disbelief $MD$: increase in disbelief.
• Certainty Factor: $CF(H,E) = MB(H,E) - MD(H,E)$

Range: $CF \in [-1, +1]$:

• $+1$ = definitely true, $-1$ = definitely false, $0$ = no information / unknown.
• Often scaled to [-100, +100] or [0,1] in systems.

Example: $CF(fever \Rightarrow flu) = 0.6$ means "fever somewhat confirms flu".

3. How It Propagates (overview — combining evidence)

• Rule with CF: IF $E$ with $CF(E)$ THEN $H$ with $CF(rule)$ → $CF(H) = CF(rule) \cdot \max(0, CF(E))$ (one common form; variants exist).
• Combine two rules concluding same $H$:
  • If both positive: $CF_{comb}=CF1+CF2-CF1\cdot CF2$
  • If both negative: $CF_{comb}=CF1+CF2+CF1\cdot CF2$
  • If opposite signs: $CF_{comb}=\frac{CF1+CF2}{1-\min(|CF1|,|CF2|)}$
• Common pitfall: combining via these formulas is ad-hoc, not probabilistically sound — but computationally cheap and modular (adding a rule just composes CFs).

4. CF vs Probability (exam contrast)

	Certainty Factor	Probability (Bayes)
Semantics	Heuristic degree of belief/disbelief, no formal sample space	Formal $P(H|E) \in [0,1]$ with Dutch-book coherence
Modularity	Modular — each rule's CF independent	Not modular — $P(H|E1,E2)$ needs joint, Bayes nets handle it
Calibration	Often tuned by experts, not data	Learned from data / coherent inference
Today	Historical — replaced by Bayes nets, but still asked	Modern standard

5. Why Still Taught

• Shows evolution: from logic → CF → probability and why probability won.
• Some expert systems still report CF-like scores ("confidence 0.85").

6. One-Line Summary

Certainty factors quantify belief as $CF=MB-MD \in [-1,1]$ with heuristic combination rules, offering modular rule-based uncertainty that is computationally simple but lacks the formal coherence of probability and Bayesian networks.

7. Quick Self-Check

1. What does $CF=-0.8$ mean?
2. Why is $CF1+CF2-CF1\cdot CF2$ less than 1 even if both are 0.8?