Abstract
The field of Artificial Intelligence is frequently hindered by a cycle where specialized knowledge representation (KR) languages, designed for specific domains, become obsolete as applications evolve. RLL (Representation Language Language) is a frame-based meta-language designed to address this by encoding the components of KR languages themselves—such as slots and inheritance mechanisms—as modifiable frames. By treating the logic of the language as data, RLL provides a highly configurable environment that can adapt to diverse representational needs, allowing researchers to "dial in" features for specific AI tasks.
The Problem: The "Bleak Cycle" of KR Design
In the early 1980s, the AI landscape was fragmented by specialized KR languages. Typically, a researcher would build a language for a specific domain, such as molecular biology or natural language processing. However, as projects matured or shifted focus, the language's original features often transformed into undesirable limitations. The common remedy—scrapping the old language to build a new one from scratch—led to a fragmented ecosystem of incompatible schemes. RLL proposes a meta-language approach: a system whose domain is the field of representational languages themselves.
The RLL Architecture
RLL is a frame-based system built on LISP that employs several novel architectural strategies to achieve flexibility:
- The "Stop Organ" Analogy: RLL is designed like a pipe organ. Each "stop" represents a pre-fabricated representational component, such as a specific inheritance type or data format. Users can "pull out" or "push in" these stops to configure the environment to mimic existing languages like KRL or FRL.
- Total Self-Description: Unlike systems where core primitives are opaque code, RLL represents its internal functions as "Units" (frames). Modifying the Unit for a function like
Getimmediately alters the behavior of the entire environment. - Slot-Defining Language: RLL provides a high-level declarative language for creating new slots. Instead of writing complex procedures, a user can define a slot through composition or other high-level descriptors.
- Truth Maintenance: Because the system is self-descriptive, it can perform "retroactive updates." If a user changes a definition, RLL ensures the internal consistency of the knowledge base.
Key Insights and Ideas
Reification of Process
The most powerful concept in RLL is the "reification" of procedural information. By turning the logic of the language into data (frames), the language becomes inspectable, editable, and capable of reasoning about itself.
The "Onion Field" vs. "Onion Skins"
Traditional software is often built in strict layers (like onion skins). RLL proposes an "onion field" where features—such as automatic inverse links—can be individually enabled or disabled in any combination, rather than in a strict linear sequence.
Cognitive Economy and Stabilizing Primitives
RLL includes mechanisms to decide when to cache a value versus when to recompute it, treating computational resources as a domain for heuristic reasoning. Furthermore, the authors noted an empirical trend: while the first 200 slots defined required 13 new "slot-combiners," the next 200 required only 5 more. This suggests that a small set of meta-primitives can cover a vast range of representational needs.
Results and Versatility
RLL has demonstrated significant functional versatility, particularly in handling complex epistemological tasks. For instance, it can represent "John’s belief about Mary’s age" as a first-class object, distinct from other historical or context-specific values for the same slot.
At the time of its development, RLL was already powering significant AI projects, including EURISKO (for heuristic discovery), VLSI (for circuit layout), and WHEEZE (for pulmonary function diagnosis). Its extensibility allows users to define non-standard inheritance modes by simply editing high-level descriptions, which RLL then translates into low-level system changes.
Limitations and Future Directions
Despite its strengths, RLL faces several challenges:
- Simulation Complexity: While the "stop organ" theory suggests RLL can mimic other languages, fully implementing these simulations remains a significant task.
- Intensional Objects: The system still struggles with representing the concept of an object versus the object itself.
- Ambiguous Semantics: Stating the exact semantics of a function (like
OR) is difficult because different users may expect different behaviors, such as sequential versus parallel evaluation. - Disambiguation: A fully automated solution for disambiguating vague user input has not yet been realized, though RLL can handle distinct representations of similar concepts (e.g., a "Typical Elephant" vs. the "Elephant Species").
Summary updated