Introduction
Classical logic operates on a binary principle: statements are either true or false. However, human communication is inherently "fuzzy." When a user describes a computer as "fast" or "relatively cheap," they are not referring to a hard numerical cutoff but rather a range of possibilities. Bridging the gap between this lexical imprecision and formal computational systems is a significant challenge in artificial intelligence.
This paper introduces Khayyam, a declarative programming language designed to handle fuzzy constraints within a first-order predicate calculus (FOPC) framework. By establishing a formal equivalence between fuzzy constraint networks and fuzzy logic models, the authors provide a robust tool for building expert systems that can process natural language requirements.
The Khayyam Language and Theoretical Foundation
The core contribution of this work is the proof that finding consistent values in a fuzzy constraint network is semantically equivalent to finding a model in fuzzy logic. This theoretical bridge allows Khayyam to support the full expressive power of FOPC while incorporating fuzzy definitions.
Key features of Khayyam include:
- Linguistic Hedges: The language supports modifiers like "very" or "fairly." Mathematically, these act as operators on membership functions (e.g., "very" might square the membership degree), allowing for more nuanced queries.
- De-Skolemization: To maintain computational feasibility, Khayyam converts fuzzy functions into function-free fuzzy predicates.
- Fuzzy Compound Propagation (FCP): The runtime system utilizes FCP, an algorithm that combines local propagation, fuzzy arc consistency, and general path consistency to refine constraints across the network.
Case Study: Computer Selection Expert System
To evaluate the practical utility of Khayyam, the authors developed an expert system for computer selection. The system operates on a set of specifications for six hypothetical computer models.
As shown in Table 1, the system ingests data from four external files: PRICES, DISKCPTY, NUMSLOTS, and SPEED. These "crisp" values serve as the ground truth against which fuzzy queries are evaluated.
Results and Fuzzy Filtering
When a user inputs a query such as fast(reqd_speed) and cheap(price), the system does not simply return a list of matches. Instead, it uses piecewise quadratic functions to assign membership degrees to each model based on how well they satisfy the imprecise criteria.
Table 2 illustrates the resulting fuzzy sets for different model selections. In a specific test case combining speed and price requirements, the system identified model3 as the most viable candidate. However, it assigned a truth-value of only 0.125. This low score indicates that while model3 is the best available option, it only partially satisfies both the "fast" and "cheap" requirements simultaneously—a common real-world scenario where trade-offs are necessary.
Conclusion
The Khayyam framework prioritizes expressive power over inferential completeness. While the FCP algorithm may not always find the most precise answer possible (it is sound but not complete), it provides a terminating and mathematically grounded method for handling complex, vague requirements. This approach is particularly valuable for expert systems where the ability to interpret natural language is more critical than exhaustive search.
Summary updated