Can Evidence be Combined in the Dempster-Shafer Theory?

John YenUAI 1987

TL;DR

This paper proposes a new relational model for the Dempster-Shafer (D-S) theory where D-S masses are treated as conditional granular distributions. This approach demonstrates that Zadeh's conjecture on the non-combinability of evidence, which arises from his unconditional relational model, does not limit the applicability of Dempster's rule in this new framework.

What Problem Does This Paper Address?

The Dempster-Shafer (D-S) theory is a powerful framework for reasoning under uncertainty, but its rule of combination, Dempster's rule, has faced challenges. A significant issue, raised by Lotfi Zadeh, concerns the "combinability" of evidence. Zadeh proposed a relational model for D-S theory where masses are represented as "granular distributions" (summaries of attribute values in a relation). From this model, he conjectured that two pieces of evidence (represented as granular distributions) cannot be combined if they lack a "conflict-free parent relation," meaning there's no underlying relation that can satisfy both. This conjecture, if true, would severely limit the practical application of Dempster's rule. This paper aims to resolve this by presenting an alternative relational model that circumvents Zadeh's combinability problem.

Methods

The authors introduce a novel relational model for the D-S theory. The core innovation is representing D-S masses as conditional granular distributions. Unlike Zadeh's model, where granular distributions are unconditional summaries, here they are explicitly conditioned on their evidential sources. This conditioning is achieved by treating the evidential sources themselves as attributes within the relational model.

The paper then demonstrates how combining two such conditional granular distributions naturally leads to a new conditional granular distribution that is conditioned on both original sources. This is achieved through a process analogous to Dempster's rule. Crucially, the paper proves that in this conditional framework, the parent relations are "partially filled," allowing for more freedom in combination. This leads to the conclusion that conditional granular distributions are always combinable whenever Dempster's rule is applicable, effectively sidestepping Zadeh's combinability problem.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Conditional vs. Unconditional Interpretation: The paper highlights that the combinability problem arises from interpreting D-S masses as unconditional summaries. By treating them as conditional summaries, the problem is resolved.
  • Evidential Sources as Attributes: Representing evidential sources as attributes in the relational model is a key mechanism for achieving the conditional interpretation.
  • "Partial" Parent Relations: The concept of "partially filled" parent relations in the conditional model is crucial. It allows for more flexibility in combining evidence compared to Zadeh's model, where parent relations are more constrained.
  • Zadeh's Conjecture Re-contextualized: The paper shows that Zadeh's conjecture, while valid for his unconditional model, does not apply to the conditional model presented here. This means Dempster's rule remains applicable in a broader range of scenarios within this new framework.

Key Results

The paper's primary result is the demonstration that their proposed relational model, using conditional granular distributions, makes evidence combinable via Dempster's rule, even in cases where Zadeh's conjecture would suggest otherwise. They prove that conditional granular distributions are always combinable as long as Dempster's rule itself is applicable (i.e., the normalization factor is not zero, which indicates total conflict).

Limitations and Future Work

  • Assumptions of Dempster's Rule: The paper acknowledges that while their model makes conditional granular distributions combinable, the application of Dempster's rule still relies on certain assumptions, such as conditional independence. The paper briefly mentions that these assumptions for a modified Dempster rule have been discussed elsewhere.
  • Scope of the Paper: The authors state that a detailed discussion of the assumptions employed in the original Dempster's rule is beyond the scope of this paper.
  • Future Work: The paper suggests that relational models offer a simple view of D-S theory and can highlight issues like combinability. The implication is that further exploration of relational models could lead to deeper understanding and practical applications of D-S theory.

Summary updated

BibTeX
@inproceedings{yen1987can,
  author = {John Yen},
  title = {{Can Evidence be Combined in the Dempster-Shafer Theory?}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1987},
}
Figure