Comprehensibility Improvement of Tabular Knowledge Bases

Atsushi Sugiura, Maximilian Riesenhuber, Yoshiyuki KosekiAAAI 1993

Abstract

Knowledge bases (KBs) in expert systems must be both logically sound and human-interpretable to ensure maintainability. Traditional logic minimization often ignores cognitive load, leading to technically correct but confusing decision tables. This paper presents an algorithm that reorganizes tabular KBs using three criteria—similarity, continuity, and conformity—to enhance comprehensibility without sacrificing logical integrity. Evaluation across real-world and artificial datasets demonstrates that the algorithm can produce human-level table organizations efficiently.

Introduction

In the development of expert systems, the "comprehensibility gap" represents a significant hurdle. While a knowledge base may be technically accurate, its utility is severely limited if human engineers cannot easily validate or update its logic. Traditionally, logic optimization has focused solely on simplicity—minimizing the number of rules or conditions. However, two logically identical sets of rules can differ vastly in their readability. For example, a rule might include redundant or cognitively jarring conditions that, while logically "safe," obscure the underlying intent. This research shifts the focus from mere minimization to meaningful organization.

Methods

The proposed approach utilizes a modified logic minimization framework based on the MINI algorithm. The core innovation lies in using specific heuristics to determine the optimal attribute order ($\sigma$) during the minimization process. This transformation is guided by three primary criteria:

  1. Similarity: Rules are structured to share similar "shapes" (cube structures), allowing humans to quickly identify common patterns.
  2. Continuity: For ordinal attributes (e.g., age, income), the algorithm groups values into continuous ranges rather than fragmented sets.
  3. Conformity: The logic is organized to reflect real-world dependencies, ensuring that "precondition" attributes are evaluated before "constrained" ones.

The algorithm pre-scans the decision table to calculate "merging ability" scores for attributes, then applies "disjoint sharp," "Expansion," and "Reduction" operations to reshape the logic into a more intuitive format.

Case Study: Exam Results

The impact of logical organization is best illustrated by comparing a poorly structured table with one optimized for comprehensibility. Consider a simple "Pass/Fail" logic for an exam and a subsequent makeup exam.

A poorly organized decision table where conditions are fragmented across rows, obscuring the primary pass criteria.

In Figure 4, the "Pass" conditions are split in a way that makes the primary rule—passing the main exam—hard to see. The first row suggests passing depends on failing the makeup, which is counter-intuitive and cognitively taxing to parse.

An optimized, highly comprehensible version of the same logic, grouping conditions by their natural real-world dependencies.

In contrast, Figure 6 presents the same logic in a way that matches human reasoning: first, it identifies that a score $\ge 60$ on the main exam is a sufficient condition for passing. Only if that condition is not met ($< 60$) does the makeup exam score become relevant. This follows the "conformity" principle by respecting the temporal and logical order of the attributes.

Results and Discussion

The algorithm was evaluated against 11 real-world and 24 artificial knowledge bases. Key findings include:

  • Human-Level Performance: In 64% of real-world cases, the algorithm's output was identical to manual organizations performed by human experts.
  • Scalability: The system successfully processed a large-scale "grinder selection" KB (158 concepts, 1,023 cubes) in just 90 seconds, a task that would take humans days.
  • Minimal Overhead: Readability improvements did not significantly increase table size; in most cases, the optimized tables were the same size as those produced by standard minimization algorithms.

While the current algorithm is limited to mutually disjoint rules, it provides a robust foundation for automated knowledge engineering. Future work will explore non-disjoint structures and more sophisticated tie-breaking heuristics to further align machine logic with human intuition.

Summary updated

BibTeX
@inproceedings{sugiura1993comprehensibility,
  author = {Atsushi Sugiura and Maximilian Riesenhuber and Yoshiyuki Koseki},
  title = {{Comprehensibility Improvement of Tabular Knowledge Bases}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}

Loading related content...

Figure