How to Represent Evidence - Aspects of Uncertain Reasoning

Claus-Rainer RollingerIJCAI 1983

Abstract

Traditional logic and one-dimensional probabilistic models often struggle to represent the nuances of human "common sense," particularly the distinction between a lack of information and the presence of conflicting information. This paper proposes a two-dimensional "Evidence Space" that tracks positive and negative evidence independently. By mapping propositions to coordinates within a unit square, the system can explicitly represent states of ignorance, certainty, and contradiction, allowing for more flexible linguistic modeling and inference.

Introduction: The Limits of One-Dimensional Truth

Traditional binary logic and even multi-valued or fuzzy logic systems typically rely on a single dimension to represent truth values. However, a value of 0.5 on a 1D scale is ambiguous: it could represent a total lack of evidence or a perfect balance of strong contradictory evidence. In human reasoning, "I don't know" is fundamentally different from "I have conflicting reports."

Claus-Rainer Rollinger argues that this ambiguity hinders AI's ability to perform common-sense reasoning. To address this, the paper introduces a representation schema that treats positive and negative evidence as independent variables, creating a richer space for handling vague and uncertain knowledge.

The Evidence Space Model

The core of the proposed methodology is the Evidence Space, defined as a unit square $[0,1] \times [0,1]$. The x-axis represents positive evidence ($EP+$), while the y-axis represents negative evidence ($EP-$).

The 2D Evidence Space mapping positive and negative evidence.

Within this space, specific coordinates correspond to logical states:

  • (1,0): Certainly true.
  • (0,1): Certainly false.
  • (0,0): Unknown (ignorance).
  • (1,1): Contradictory.

Reasoning in this system is geometric. Every proposition is assigned an Evidence Point (EP). Verification is not a binary match but a matter of proximity. A premise defines a Goal Point (GP)—the ideal evidence state—and a Neighborhood, which is a sphere of tolerance around that point. If the current evidence falls within this neighborhood, the premise is considered compatible. This allows the system to handle "relevance" by adjusting the size of the neighborhood; less relevant premises have larger neighborhoods, making them harder to falsify.

Linguistic Mapping and Inference

The framework treats natural language modifiers as mathematical functions that transform coordinates in the Evidence Space. Phrases like "It is the case" or "It may be true" act as operators that move an Evidence Point toward specific regions of the square.

Mapping natural language statements to evidence coordinates.

By modeling nested modifiers (e.g., "It is false that it may be true...") as sequential transformations, the system can compute complex evidence states. For example, a statement might result in a coordinate like $(0.7, 1)$, indicating a state of high conflict where negative evidence is certain but positive evidence remains significant.

Conclusion and Future Work

The two-dimensional evidence space provides a powerful qualitative tool for AI reasoning, successfully separating ignorance from contradiction. However, several areas remain for future development. The exact vector calculus for computing conclusion evidence from premise quality is still under investigation. Furthermore, the relationship between these 2D coordinates and traditional truth values needs formalization. Future iterations of the PROJECT KIT aim to integrate uncertainty within the inference rules themselves into this same geometric schema.

Summary updated

BibTeX
@inproceedings{rollinger1983how,
  author = {Claus-Rainer Rollinger},
  title = {{How to Represent Evidence - Aspects of Uncertain Reasoning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure