Dempster-Shafer Theory

The foundational set of all mutually exclusive and exhaustive hypotheses, denoted Θ (Theta). For a sensor fusion problem, this might be {pedestrian, cyclist, vehicle, static_object}. DST operates on the power set 2^Θ, which includes all subsets (e.g., {pedestrian, cyclist}), enabling the model to express uncertainty between specific alternatives without committing to a precise probability distribution.

Also called a mass function (m) , this is the core evidence carrier. It assigns a value between 0 and 1 to every subset of the frame of discernment, where m(∅) = 0 and the sum of all masses equals 1. Critically, m({pedestrian, cyclist}) = 0.3 explicitly states that 30% of the evidence supports the object being either a pedestrian or a cyclist, without specifying which—a direct representation of epistemic uncertainty impossible in standard probability.

DST defines an evidential interval [Bel(A), Pl(A)] for every proposition A, replacing the single point-value of probability.

• Belief (Bel): The total mass necessarily supporting A. Bel(A) = Σ m(B) for all subsets B ⊆ A. It is the lower bound of probability.
• Plausibility (Pl): The total mass that does not contradict A. Pl(A) = Σ m(B) for all subsets B ∩ A ≠ ∅. It is the upper bound. The gap Pl(A) - Bel(A) quantifies ignorance about A.

The conjunctive fusion operator used to combine independent sources of evidence (e.g., a camera BBA and a radar BBA). For two mass functions m1 and m2, the combined mass for hypothesis A is calculated as the sum of the products of masses for all pairs of subsets whose intersection is A, normalized by 1 - K, where K is the conflict coefficient. This normalization step is the source of DST's ability to handle conflicting sensor data, though it can produce counter-intuitive results in high-conflict scenarios.

The conflict coefficient (K) measures the degree of disagreement between evidence sources. When K is high, Dempster's rule can yield paradoxical results (the Zadeh paradox). To address this, alternative combination rules have been developed:

• Yager's Rule: Assigns conflict mass to the universal set Θ, treating conflict as ignorance.
• Dubois-Prade Rule: Distributes conflict mass to the union of the conflicting propositions.
• PCR6 (Proportional Conflict Redistribution): Redistributes conflicting mass proportionally to the sources involved, preserving specificity in high-conflict fusion.

DST provides an evidential interval, but a final decision often requires a single probability. The pignistic transformation converts a mass function into a probability distribution for decision-making by equally distributing the mass of each non-singleton subset among its elements. For m({A, B}) = 0.3, each of A and B receives 0.15. This step bridges the gap between the credal level (evidence combination) and the pignistic level (action).

The classical probabilistic framework that Dempster-Shafer generalizes. Bayesian inference requires precise prior probabilities and represents uncertainty as a single additive measure, where belief in a hypothesis forces belief in its negation. Unlike DST, it cannot represent ignorance—the inability to distinguish between lack of evidence and negative evidence. In sensor fusion, Bayesian methods struggle when sensor models are incomplete or when expressing 'I don't know' is critical for safe operation.

The lower bound of evidential support in DST, denoted as Bel(A). It represents the total mass of evidence that directly supports proposition A, summing all basic probability assignments for subsets contained within A. In manufacturing sensor fusion, a belief function quantifies the minimum certainty that a defect exists based on current evidence, providing a conservative estimate that avoids overconfident decisions when data is sparse or contradictory.

The upper bound of evidential support, denoted as Pl(A). It represents the maximum possible belief that could be assigned to proposition A if all unknown evidence were resolved in its favor. Calculated as 1 - Bel(¬A). The interval [Bel(A), Pl(A)] defines the evidential uncertainty range. In industrial monitoring, a wide belief-plausibility gap signals epistemic uncertainty—indicating that more sensor data is needed before committing to an automated decision.

The core fusion operator in DST that combines independent bodies of evidence into a single belief structure. It uses orthogonal sum to conjunctively merge mass functions while normalizing away conflict. Key properties:

• Associative and commutative: fusion order does not matter
• Conflict-sensitive: high conflict produces low combined belief
• Criticism: Zadeh's counterexample shows counterintuitive results under extreme conflict, motivating alternative rules like Yager's and Dubois-Prade's

A two-level reinterpretation of DST by Philippe Smets that separates belief representation from decision-making. The credal level handles evidence combination using the conjunctive rule without normalization, preserving conflict as a measure of inconsistency. The pignistic level transforms belief functions into probabilities only when a decision is required, using the pignistic transformation: BetP(ω) = Σ m(A)/|A| for all A containing ω. This avoids premature probabilistic commitment.

The fundamental unit of evidence in DST, also called a mass function m: 2^Ω → [0,1]. A BPA distributes evidential weight across the power set of the frame of discernment Ω, satisfying:

• m(∅) = 0 (closed-world assumption)
• Σ m(A) = 1 for all A ⊆ Ω Unlike probability distributions, mass can be assigned to compound hypotheses like {A, B}, explicitly modeling ambiguity when sensors cannot distinguish between failure modes.