Argumentation as a General Framework for Uncertain Reasoning

John Fox, Paul J. Krause, Morten Elvang-GøranssonUAI 1993

TL;DR

This paper proposes "argumentation" as a unified, general framework for handling various forms of uncertain reasoning. By extending traditional logic with labeled deductive systems, it allows for the construction and evaluation of arguments, providing a principled way to accommodate diverse uncertainty calculi, from probabilistic to qualitative and defeasible approaches.

What Problem Does This Paper Address?

The paper tackles the challenge of unifying the diverse landscape of uncertainty reasoning methods used in AI and other fields. While Bayesian probability theory is dominant, its universal applicability is questioned, leading to the development of numerous alternative schemes like fuzzy sets, possibility theory, and belief functions. Furthermore, there's a growing debate about whether uncertainty must be represented quantitatively, with arguments for qualitative and logical methods that better capture intuitions about doubt and belief, especially when quantitative data is scarce. The authors aim to bridge this gap by providing a formal framework that can accommodate and integrate these different approaches to uncertainty management.

Methods

The core of the paper is the development of a labeled deductive system for argumentation. In this system, propositions are not just asserted but are "doubly labeled" with the grounds on which they are based and a representation of confidence attached to the argument. Argument construction is formalized using a generalized argument consequence relation based on the ∧, →-fragment of minimal logic. This allows for the creation of arguments that not only prove propositions but also indicate support for or doubt in them.

Confidence labels can be drawn from various "dictionaries" (e.g., discrete symbols like +, ++, or numerical intervals like [0,1]). The paper introduces the concept of flattening functions to aggregate arguments and their associated confidence labels. These functions can combine multiple arguments for or against a proposition into a single confidence value. The authors also present an Argumentation Theorem Prover (ATP) that can be instantiated with different dictionaries and aggregation functions, demonstrating its flexibility. They define several abstract dictionaries and corresponding aggregation criteria (e.g., Dgeneric, Dbounded, Ddelta, Dbounded-delta) to capture different aspects of uncertainty reasoning, including defeasibility and conflict resolution.

Datasets

Not discussed in the paper. The paper focuses on a formal framework and uses illustrative examples with hypothetical databases rather than specific, standardized datasets.

Key Insights and Ideas

  • Argumentation as a Unifying Principle: The paper posits that argumentation, viewed as the process of constructing and evaluating reasons for or against propositions, provides a common logical structure underlying various uncertainty calculi.
  • Doubly Labeled Deductive Systems: The novel approach of labeling propositions with both grounds and confidence representations is key to capturing the structure and strength of arguments.
  • Flexibility of Confidence Labels: The framework is designed to be agnostic to the specific type of confidence representation, allowing for integration of symbolic, qualitative, and quantitative measures.
  • Formalizing Defeasibility: The framework naturally accommodates defeasible reasoning by allowing arguments to be "defeated" (discounted or rebutted) by other arguments, a crucial aspect of real-world reasoning.
  • Aggregation as a Core Operation: The paper emphasizes the importance of aggregation functions for combining evidence from multiple arguments to arrive at an overall confidence.

Key Results

The paper presents a formal system (LA) and several abstract dictionaries and aggregation criteria that demonstrate how diverse uncertainty reasoning approaches can be unified under the argumentation framework. The examples illustrate how the framework can handle:

  • Qualitative reasoning: Using discrete confidence labels like + and ++ (e.g., the Dbounded example).
  • Quantitative reasoning: Propagating numerical confidence coefficients and aggregating them using rules similar to Bernoulli's rule (e.g., the Dnum example).
  • Defeasible reasoning: By introducing concepts of argument defeat (discounting and rebutting) and selective aggregation based on undefeated arguments.

The authors claim their framework offers a general basis for developing reasoning techniques that accommodate quantitative, qualitative, and wider logical approaches to uncertainty management.

Limitations and Future Work

  • Informal Presentation: The authors acknowledge that much of the discussion is kept relatively informal for brevity, with rigorous development detailed in cited works.
  • Conflict Resolution: While the paper introduces concepts for resolving conflicts (defeating arguments), a more extensive account of conflict resolution strategies is mentioned as being found elsewhere.
  • Practical Implementation: The paper focuses on the formal framework and theoretical underpinnings. While it mentions practical use in medical decision-making, further work on extending practical techniques and applications is implied.
  • Meta-Argumentation: The paper notes that concepts previously termed "Meta-argumentation" (where arguments are objects of arguments) have been integrated into flattening functions for a more coherent presentation. This suggests a potential for further exploration of meta-level reasoning.
  • Dependencies in Aggregation: The Dnum example briefly mentions that a more elaborate version of the example considering dependencies between arguments is given in other work, implying this is an area for further development.

Summary updated

BibTeX
@inproceedings{fox1993argumentation,
  author = {John Fox and Paul J. Krause and Morten Elvang-Gøransson},
  title = {{Argumentation as a General Framework for Uncertain Reasoning}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure