Skip to content
Back to search results

Dempster-Shafer Theory of Evidence

Introduced by
Arthur P. Dempster & Glenn Shafer
Year
1976
Type
Uncertainty calculus for combining evidence
Family
Evidence theory
Font size

What does it do?

Dempster-Shafer theory is a mathematical framework for reasoning under uncertainty that generalizes Bayesian probability by representing ignorance explicitly.

Instead of forcing a single probability on each hypothesis, it assigns belief mass to sets of hypotheses and derives a belief-plausibility interval, and it provides Dempster's rule for fusing evidence from multiple independent sources. Developed from Arthur Dempster's 1967 work and Glenn Shafer's 1976 monograph, it underpins evidential reasoning and sensor/decision fusion.

The intuition behind the method

Bayesian probability insists you split your belief fully across the hypotheses, even when you simply do not know — you must put a number on each. Dempster-Shafer lets you say 'I have evidence for A-or-B but cannot yet split it', by assigning mass to the set {A, B} rather than to A and B individually. From these masses it computes two bounds for any hypothesis: belief (the support that necessarily commits to it) and plausibility (the support that does not rule it out). The gap between them measures your ignorance. When a second source of evidence arrives, Dempster's rule combines the two by intersecting their focal sets and renormalizing away the conflict.

History

Arthur Dempster introduced upper and lower probabilities induced by multivalued mappings in 1967, and Glenn Shafer recast and extended this into a self-contained theory of evidence in his 1976 book A Mathematical Theory of Evidence, giving belief functions, plausibility, and the rule of combination. The framework spread through artificial intelligence and information fusion as 'evidential reasoning', and later refinements — the transferable belief model (Smets), alternative combination rules for conflict, and the evidential reasoning approach in MCDM — addressed its known difficulties while preserving its explicit treatment of ignorance.

When to use it

Use Dempster-Shafer theory when you must combine uncertain, partial, or conflicting evidence from multiple sources and want to represent ignorance explicitly rather than assume a full prior — typical in sensor fusion, multi-source classification, diagnostic and target-recognition systems, risk and reliability assessment, and GIS evidence overlay. It is valuable when assigning precise Bayesian priors is artificial and when distinguishing 'no evidence' from 'evidence of no' matters. Caveats: Dempster's rule can produce counter-intuitive results under highly conflicting sources (Zadeh's paradox), computation grows with the power set of hypotheses, and the independence assumption between sources is often only approximate. Alternatives/extensions include Bayesian networks, the transferable belief model, and other conflict-tolerant combination rules.

Limitations and considerations

  • Dempster's rule can give counter-intuitive results under high conflict (Zadeh's paradox).
  • Computation scales with the power set of the frame, costly for many hypotheses.
  • Assumes sources are independent, which is often only approximately true.
  • Eliciting meaningful mass functions can be difficult in practice.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Represents ignorance explicitly via mass on sets, not forced single probabilities.
  • Provides belief and plausibility bounds that quantify how much is unknown.
  • Offers a principled rule for fusing evidence from multiple independent sources.
  • Generalizes Bayesian probability, recovering it when all mass is on singletons.

Applications

  • Multi-sensor and multi-source data fusion for classification and target recognition.
  • Diagnostic and decision-support systems combining uncertain expert/test evidence.
  • Risk, reliability, and safety assessment with incomplete information.
  • GIS evidence overlay and suitability mapping from heterogeneous layers.
  • Evidential reasoning approaches in multi-criteria decision analysis.

How it works

How it works

1

Evidence is encoded as a basic probability assignment (mass function) m over subsets of the frame of discernment Θ (the set of mutually exclusive hypotheses). Mass can be placed on any subset — including the whole frame Θ, which represents complete ignorance — and the masses sum to one, with the empty set getting zero.

Formula
m: 2^{\Theta} \to [0,1],\quad m(\emptyset)=0,\ \sum_{A \subseteq \Theta} m(A) = 1
2

From the masses, two measures are derived for any hypothesis A: belief Bel(A) sums the mass of all subsets entirely inside A (the evidence that fully commits to A), while plausibility Pl(A) sums the mass of every subset that intersects A (the evidence that is at least compatible with A). Together they form the interval [Bel(A), Pl(A)], whose width quantifies remaining uncertainty.

Formula
\text{Bel}(A) = \sum_{B \subseteq A} m(B),\quad \text{Pl}(A) = \sum_{B \cap A \neq \emptyset} m(B)
3

Dempster's rule of combination fuses two independent mass functions: for each pair of focal sets it multiplies their masses and assigns the product to their intersection, then normalizes by 1 − K, where K is the total mass that fell on the empty set (the conflict between sources). The result is a new mass function summarizing the pooled evidence.

Formula
m_{12}(A) = \dfrac{1}{1-K}\sum_{B \cap C = A} m_1(B)\,m_2(C),\quad K = \sum_{B \cap C = \emptyset} m_1(B)\,m_2(C)
4

Decisions use the belief-plausibility interval — for example, choosing the hypothesis with the highest belief, the highest plausibility, or the highest pignistic (probability-like) transform of the masses. The explicit interval lets a system defer or request more evidence when belief and plausibility diverge widely.

Formula
[\text{Bel}(A),\ \text{Pl}(A)]\ \text{= belief interval for } A

Common pitfalls

  • Applying Dempster's rule to highly conflicting sources without checking the conflict mass K.
  • Confusing belief with probability — belief is a lower bound, not a point estimate.
  • Ignoring the combinatorial blow-up when the frame of discernment is large.
  • Combining dependent sources as if independent, double-counting shared evidence.

Frequently asked questions

How does Dempster-Shafer differ from Bayesian probability?

Bayesian probability assigns a single number to each hypothesis and requires a full prior. Dempster-Shafer assigns mass to sets of hypotheses, letting it represent ignorance directly, and summarizes each hypothesis by a belief-plausibility interval. It generalizes the Bayesian approach, which is recovered when all mass sits on individual hypotheses.

What are belief and plausibility?

Belief Bel(A) is the total evidence that fully commits to A (a lower bound on its probability); plausibility Pl(A) is the total evidence merely compatible with A (an upper bound). The interval [Bel(A), Pl(A)] brackets the true support for A, and its width reflects how much remains unknown.

What is the conflict problem in Dempster's rule?

When two sources strongly disagree, most of the combined mass falls on the empty set (the conflict K). Normalizing it away can yield counter-intuitive results — Zadeh's classic paradox. High conflict signals that Dempster's rule may be inappropriate, motivating alternative combination rules or a review of source reliability.

Related methods

Sources

  1. Dempster, A. P. (1967). Upper and lower probabilities induced by a multivalued mapping. The Annals of Mathematical Statistics, 38(2), 325–339.

    Open DOI reference
    Copied.
  2. Shafer, G. (1976). A Mathematical Theory of Evidence. Princeton University Press.

    ISBN: 978-0-691-08175-5
    Copied.