A Reasoning Model Based on an Extended Dempster-Shafer Theory

John YenAAAI 1986

TL;DR

This paper introduces a novel reasoning model that extends the Dempster-Shafer (D-S) theory to better handle uncertain relationships between evidence and hypotheses. By transforming the D-S theory's multi-valued mapping into a probabilistic one and using "basic certainty values" to represent belief updates, the model provides probability bounds consistent with Bayesian theory while retaining the D-S theory's ability to express ignorance.

What Problem Does This Paper Address?

Expert systems often deal with situations where evidence supports a group of mutually exclusive hypotheses, but doesn't fully differentiate between them. For instance, a medical symptom might suggest "rheumatoid arthritis" but not distinguish between its "seropositive" and "seronegative" subtypes. Traditional Bayesian approaches struggle here because they force a precise probability distribution, losing information about uncertainty. The Dempster-Shafer (D-S) theory offers a way to represent ignorance but faces challenges in formally expressing uncertain relationships between evidence and hypotheses, leading to belief functions that aren't always probability bounds. This paper aims to bridge this gap by developing a D-S-based model that can represent these uncertain associations and produce meaningful probability bounds.

Methods

The core of the proposed model lies in two key extensions to the Dempster-Shafer theory:

  • Probabilistic Multi-set Mapping: Instead of a strict one-to-one or one-to-many mapping with certainty (probability of 0 or 1), the paper introduces a probabilistic multi-set mapping. This allows an evidential element to be associated with a set of hypothesis groups, each with a conditional probability. This directly addresses the problem of representing uncertain relationships like "there's an 80% chance of hypothesis A given evidence e."
  • Basic Certainty Values (BCVs): To address the issue of Dempster's rule potentially over-weighting prior probabilities when combining evidence, the authors introduce "basic certainty values." These values are derived by discounting the prior probability from the mass distribution, effectively measuring "belief update" rather than absolute belief. Dempster's rule is then applied to these BCVs, and the combined BCV is transformed back into a combined belief function. This approach ensures that the resulting belief and plausibility intervals act as lower and upper probability bounds under specific conditional independence assumptions.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and provides a hypothetical example for illustration.

Key Insights and Ideas

  • Bridging Bayesian and D-S: The model successfully reconciles the strengths of both Bayesian and D-S theories. It adopts the D-S ability to represent ignorance while ensuring that the resulting probability bounds are consistent with Bayesian principles under certain independence conditions.
  • Belief Update vs. Absolute Belief: The introduction of "basic certainty values" is a crucial insight. It distinguishes between the absolute degree of belief (represented by the standard D-S mass function) and the update to that belief caused by new evidence. This is a more nuanced way to handle evidence combination.
  • Weaker Independence Assumptions: The paper argues that its conditional independence assumptions are weaker than those typically used in systems like MYCIN or PROSPECTOR, potentially allowing for more robust reasoning in complex hierarchical hypothesis spaces.
  • Representation of Ignorance: A significant advantage highlighted is the model's ability to coherently express and update degrees of ignorance, a key limitation of purely Bayesian approaches.

Key Results

The paper demonstrates that the proposed model produces belief intervals that serve as lower and upper probability bounds for hypotheses, given the evidence and under specific conditional independence assumptions. The authors provide a detailed example illustrating how their model combines evidence from two sources, showing how the resulting belief intervals provide a more nuanced view than a purely Bayesian approach, which can over-allocate mass to hypotheses with high prior probabilities.

Limitations and Future Work

The authors acknowledge several areas for future research:

  • Chaining of Reasoning: Developing mechanisms to perform chains of reasoning using the proposed model.
  • Computational Complexity: Analyzing and optimizing the computational complexity of the model, which is currently dominated by Dempster's rule.
  • Interval-Based Decision Making: Further research into decision-making strategies that effectively utilize the interval probabilities generated by the model.
  • Generalization of Mapping: The paper notes that extending the probabilistic mapping to cases where the evidence space probability distribution is not precisely known presents a challenge, and further justification for the aggregation rule in such scenarios is needed.

Summary updated

BibTeX
@inproceedings{yen1986reasoning,
  author = {John Yen},
  title = {{A Reasoning Model Based on an Extended Dempster-Shafer Theory}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure