Presented at AAAI-88, this research explores the integration of expert systems with fuzzy logic. While traditional forward-chaining systems, such as OPS5, rely on "crisp" binary logic, the authors introduce FMUFL (Fuzzy Multi-Formalism Language). This language is designed to handle the inherent ambiguity of human language—terms like "cheap," "very hungry," or "fairly expensive"—while maintaining the structured control required for complex task management.
The Challenge of Fuzzy Conflict Resolution
In the late 1980s, forward-chaining production systems were the standard for AI expert systems. These systems operate by matching "if-then" rules against a working memory of facts. However, standard systems were strictly binary: a fact either matched a rule or it did not.
The real world is rarely so clear-cut. If a rule triggers when a price is "cheap," and the price is $35, fuzzy logic allows for a "membership grade" (e.g., $35 might be 0.25 "cheap"). The core challenge arises in conflict resolution: if multiple rules match with varying degrees of truth (e.g., 0.9, 0.75, 0.4), which one should the system execute? Simply selecting the "truest" rule can inadvertently disrupt the logical flow of a complex program.
FMUFL: Bridging the Gap
FMUFL was developed to bridge the gap between classical forward-chaining and fuzzy reasoning. It supports two primary types of fuzzy representation:
- Lexical Imprecision: Support for "I-type" fuzzy sets (continuous variables like price) and finite fuzzy sets (discrete lists).
- Semantic Hedges: The system handles modifiers mathematically. For example, "very" squares the membership grade to make criteria stricter, while "fairly" takes the square root to be more inclusive.
For complex rules with multiple conditions, the system utilizes the min operator to determine the overall truth value of the match, ensuring that the rule is only as true as its weakest condition.
The Sieve Approach to Conflict Resolution
The authors propose treating conflict resolution as a series of filters, or "sieves." Rules are passed through these filters until a single candidate remains. Crucially, for the system to remain task-oriented, fuzzy truth values must be integrated carefully with traditional metadata like time tags.
To preserve the programmer's intended flow of control, the authors established a specific 6-step order for these sieves:
- Absolute Truth-Value: Discard matches below a threshold (default 0.5).
- Refraction: Prevent the same rule from firing on the same data twice.
- Recency: Prefer rules matching the most recently added data (using time tags).
- Element Specificity: Prefer rules matching more specific sets of facts.
- Test Specificity: Prefer rules with more complex "if" conditions.
- Relative Truth-Value: Finally, if a tie remains, select the "truest" match.
Key Insights and Results
The most significant insight is the critique of "truth-first" filtering. If a system prioritizes the highest truth value immediately, it might fire a "task cleanup" rule (100% true) before finishing the actual task (which might only be 75% true). By placing the relative truth-value sieve at the very end (Step 6), FMUFL ensures that the intended task hierarchy is not overridden by high membership grades.
Furthermore, FMUFL maintains "upward compatibility" with classical systems. If fuzzy terms are removed, the system behaves exactly like the industry-standard OPS5, making it a practical tool for engineers transitioning from binary to fuzzy logic.
Conclusion and Future Work
While FMUFL provides a robust framework for fuzzy task management, the authors acknowledge challenges in predictability, as it can be difficult for programmers to anticipate exact truth values at runtime. Future work may explore parallel rule firing, though the current serial approach remains superior for the hierarchical, task-oriented programming required in complex expert systems.
Summary updated