Fuzzy sets & Dempster-Shafer (breadth)

Fuzzy sets: degrees of membership

Classical (*crisp*) sets are binary — an element is in or out. A fuzzy set $F$ of a universe $S$ is defined by a membership function $m_F(s) \in [0,1]$ giving each element a *degree* of membership. A 5′9″ man might belong to *medium height* with degree $0.6$ and *tall* with degree $0.4$ simultaneously — which deliberately violates two classical axioms: exclusive membership, and the law of the excluded middle (he is partly in the set *and* partly in its complement).

The motivation is Zadeh’s key distinction: vagueness is not randomness. “Tall” is *vague* — there is no sharp threshold — but a person’s height is not *random*. Probability theory handles randomness; fuzzy logic handles vagueness. Conflating the two is the classic error to avoid.

The fuzzy control cycle (recognition level)

Fuzzy controllers (famously the inverted pendulum, and now consumer electronics) run a four-step cycle:

1. Fuzzification — convert a crisp input (e.g. angle $\theta = 1$) into membership degrees across fuzzy regions (e.g. *Zero* $= 0.5$, *Positive* $= 0.5$).
2. Rule evaluation — fire all matching fuzzy rules, combining premises with AND = min and OR = max (the same algebra as certainty factors), each producing an output region clipped to the premise’s strength.
3. Aggregation — union all activated output regions.
4. Defuzzification — collapse the aggregated fuzzy region back to a single crisp control value, usually by the centroid (center-of-mass) method.

The practical win: instead of solving differential equations in real time, the controller uses intuitive linguistic rules (“IF angle is Positive AND speed is Negative THEN do nothing”) and runs at near-circuit speed.

Dempster-Shafer: belief intervals for ignorance

Dempster-Shafer (D-S) theory assigns each hypothesis not a single number but a belief interval $[bel, pl]$, where $bel$ is the *direct supporting evidence* and the plausibility $pl = 1 - bel(\lnot p)$ is an upper bound on belief. With no evidence, a hypothesis sits at $[0, 1]$ — *genuine ignorance, explicitly represented*. This is D-S’s headline advantage over Bayes: a Bayesian must assign a prior (say $0.5$), which cannot distinguish “I have no evidence” from “I have strong, perfectly balanced evidence.” The D-S interval *shrinks as evidence accumulates*, giving a natural measure of how much you actually know. Dempster’s rule of combination merges belief functions from independent sources, summing the products of masses over intersecting hypothesis sets and renormalizing by removing the mass that landed on the empty set (the conflict). A large empty-set mass signals that the sources strongly disagree.