Default Reasoning as Likelihood Reasoning

Elaine RichAAAI 1983

Abstract

This paper challenges the traditional binary view of default reasoning in artificial intelligence, proposing instead a framework based on likelihoods. By utilizing Certainty Factors (CFs), the author demonstrates how to resolve logical contradictions and manage complex inheritance hierarchies more effectively than rigid non-monotonic logics.

Introduction

In the early 1980s, AI research was preoccupied with "non-monotonic reasoning"—the challenge of retracting conclusions when faced with new evidence. The classic example is assuming a bird can fly until discovering it is an ostrich. Traditional formal logics proved too brittle for these "usually true" statements, often leading to mathematical complexity or ambiguous "multiple extensions" where the system could not determine which outcome was more probable. Elaine Rich's 1983 paper argues that defaults should not be treated as binary logical absolutes but as degrees of belief or likelihoods.

The Certainty Factor Framework

Rich proposes adopting the Certainty Factor (CF) model, originally used in the MYCIN expert system, to represent default rules. In this system, every proposition is assigned a CF between -1 (definitely false) and 1 (definitely true), where 0 represents no evidence. Default rules are expressed as implications with associated weights.

Figure 5: Representation of default rules for the bird problem using Certainty Factors.

As shown in the rules above, a general rule like "birds fly" is given a high but non-absolute CF (e.g., 0.9), while specific facts like "Sandy is an ostrich" are given a CF of 1.0. This allows for numerical propagation of belief: if a system is 80% certain an object is a bird, and the "birds fly" rule has a 0.9 confidence, the resulting confidence in flight is $0.8 \times 0.9 = 0.72$.

Resolving Conflicts in ISA Hierarchies

A major contribution of the paper is the Monotonic Consistency Constraint (MCC). To ensure that specific knowledge correctly overrides general defaults, properties attached to more specific nodes in a hierarchy must have higher absolute CFs than conflicting properties at more general nodes. This ensures that the more specific information (e.g., "ostriches don't fly") always carries more weight than the general default.

Figure 2: Semantic networks illustrating complex inheritance and potential conflicts in property assignment.

This approach handles "tangled hierarchies" (Figure 2) where an entity might inherit properties from multiple paths. Unlike "shortest path" heuristics used in early semantic networks, the likelihood approach uses explicit CFs to determine the net belief. For instance, in Figure 2(a), if Sandy is both an ostrich (typically not tame) and a pet (typically tame), the system calculates a final CF by weighing the evidence from both paths rather than stalling or choosing arbitrarily.

Key Insights and Results

Rich identifies three distinct types of knowledge that logic typically conflates:

  • Definitional: "People are mammals" (CF = 1.0).
  • Prototypical: "People can see" (CF = 0.99).
  • Statistical: "Southerners are Democrats" (CF = 0.8).

By assigning different CFs to these categories, a single system can represent them uniformly. The results of this approach include deterministic conflict resolution, providing a single weighted answer instead of multiple contradictory extensions. Furthermore, it simplifies rule bases by eliminating the need for complex "UNLESS" clauses, as the system naturally handles exceptions by weighing evidence.

Limitations and Future Work

Despite its elegance, the paper acknowledges several hurdles. The primary issue is the source of the Certainty Factors; assigning these numbers manually is subjective and difficult to scale. Furthermore, the paper notes a tension between "subtracting" negative evidence versus "overriding" it entirely, a distinction that requires further formalization. Finally, the computational cost of propagating these factors through massive knowledge bases remained an open question for future research.

Summary updated

BibTeX
@inproceedings{rich1983default,
  author = {Elaine Rich},
  title = {{Default Reasoning as Likelihood Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure