Parallel Formulation of Evidential-Reasoning Theories

TL;DR

This paper tackles the challenge of unifying diverse "evidential reasoning" (ER) approaches used in expert systems. By proposing a general four-part paradigm, it structures and compares five key ER theories (Bayes, Convex Bayes, Dempster-Shafer, Kyburg, and Possibility), revealing their underlying assumptions, similarities, and differences. This work aims to provide a clearer understanding and a foundation for selecting the most appropriate ER method for specific expert system tasks.

What Problem Does This Paper Address?

Expert systems often need to reason with uncertain or incomplete information. This is known as evidential reasoning (ER). However, there's no single, universally agreed-upon way to handle ER. Different fields like statistics and philosophy have contributed various approaches, leading to a lack of common terminology and assumptions. This makes it difficult to compare these methods, understand their strengths and weaknesses, and choose the best one for a given expert system application. The paper aims to bridge this gap by providing a structured way to analyze and compare these diverse ER theories.

Methods

The core contribution of this paper is a general four-part paradigm for structuring ER problems. This paradigm acts as a framework to analyze different ER approaches:

  1. Background Elements: Defines the domain of discourse (the world-model) and current knowledge, including potential costs of actions.
  2. Observation Reports: Describes the structure and content of incoming information about the external world.
  3. Updating Mechanism: Outlines the rules and algorithms for revising knowledge based on new observations.
  4. Decision Mechanism: Details the rules and algorithms for making choices given the updated knowledge.

Using this paradigm, the authors then systematically analyze five prominent ER approaches: classical Bayes, convex Bayes, Dempster-Shafer, Kyburg, and possibility theory. Each approach is presented by detailing how it fits into the four parts of the paradigm, allowing for a direct comparison of their assumptions, how they represent belief states (e.g., as single probabilities, intervals, convex sets, or fuzzy sets), and their updating and decision-making processes.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • A Unified Structure is Crucial: The paper demonstrates that a common paradigm can effectively highlight the similarities and differences between seemingly disparate ER theories. This is key to moving beyond ad-hoc comparisons.
  • Belief State Representation Varies Wildly: A major differentiator between approaches is how they represent uncertainty. This ranges from precise probability values (classical Bayes) to intervals (Kyburg), convex sets (convex Bayes), and fuzzy sets (Possibility theory). This choice has significant implications for how evidence is processed and decisions are made.
  • Updating Mechanisms are Diverse: While Bayes' Theorem is central to some approaches (classical and convex Bayes), others use different mechanisms (Dempster-Shafer's mass functions, Kyburg's direct inference, and Possibility theory's fuzzy logic). This impacts how quickly and how profoundly beliefs change with new evidence.
  • Decision Mechanisms Lag Behind: A significant challenge is extending decision-making frameworks (like expected loss) from point-valued beliefs to interval or fuzzy representations. This remains an open question for several approaches.
  • The Algebra of Statements is Foundational: Regardless of the specific ER approach, a well-defined algebra of statements that sharply defines the domain of discourse is a common and critical element.

Key Results

The paper doesn't present quantitative results from experiments but rather a qualitative analysis and comparison of five ER approaches. The key "results" are the structured descriptions of each approach within the proposed paradigm, which reveal:

  • Classical Bayes: Uses single probability functions and Bayes' Theorem for updating.
  • Convex Bayes: Extends classical Bayes by using convex sets of probability functions, allowing for more nuanced belief representation.
  • Dempster-Shafer: Employs mass functions to represent belief and plausibility, offering a way to combine evidence that doesn't require precise prior probabilities.
  • Kyburg: Focuses on epistemological probability (credibility relative to knowledge) and uses interval-valued probabilities, often employing direct inference.
  • Possibility Theory: Utilizes fuzzy membership functions and fuzzy decision functions, particularly useful for handling vague or imprecise information.

The paper highlights that while the structure of the "algebra of statements" is similar across approaches, the representation of belief states and the updating mechanisms differ significantly.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Lack of General Decision Mechanisms for Interval/Fuzzy Beliefs: A major open question is how to reliably choose among actions when belief states are represented by intervals or fuzzy sets. The paper notes that extending expected-loss calculations is possible but not fully resolved.
  • Dynamic Algebras: The current paradigm assumes a static algebra of statements. Future work could explore dynamic algebras that adapt to changing real-world conditions.
  • Criteria for Choosing ER Approaches: General criteria for selecting the most appropriate ER approach for different expert system tasks are still needed.
  • Computational Experiments: The authors are currently conducting computational experiments to compare the evolution of belief states under different approaches.
  • Extension to Other Theories: The parallel formulation is being extended to include other theories like certainty theory and Cohen's endorsement theory.

Ultimately, the paper aims to lay the groundwork for developing such guidelines and a deeper understanding of evidential reasoning in artificial intelligence.

Summary updated

BibTeX
@inproceedings{1985parallel,
  title = {{Parallel Formulation of Evidential-Reasoning Theories}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure