Less Than General Production System Architectures

Douglas B. Lenat, John P. McDermottIJCAI 1977

Abstract

In the late 1970s, the field of artificial intelligence faced a tension between the pursuit of general-purpose cognitive models and the practical requirements of high-performance expert systems. In their 1977 seminal work, "Less Than General Production System Architectures," Douglas Lenat and John McDermott argued that the constraints of "pure" production systems—designed for generality—hindered the development of truly expert machines. By analyzing successful systems like MYCIN and DENDRAL, the authors proposed a shift toward architectures that relax traditional constraints on data storage, rule structure, and interpreter logic to achieve greater efficiency, intelligibility, and self-modifying capabilities.

Introduction

During the mid-1970s, the dominant paradigm for AI research was the Production System (PS). Pioneered by researchers like Allen Newell, the PS was envisioned as a general-purpose architecture capable of simulating human cognition across any domain. A "pure" PS consists of three rigid components: a Working Memory (facts), a Production Memory (IF-THEN rules), and an Interpreter (the matching engine).

However, a gap emerged between theory and practice. While purists sought domain-independent generality, practitioners building the first generation of expert systems found that they had to "break the rules" of the pure PS to achieve success. Lenat and McDermott’s paper formalizes these deviations, arguing that for a system to exhibit high-level expertise, it must move away from general-purpose architectures toward "less than general" designs tailored to specific domains.

The Problem with Generality

The drive for generality in early AI came with significant architectural costs. In a pure PS, rules were required to be "fine-grained" and independent, data was ephemeral, and the interpreter was a fixed, opaque black box. This rigidity created several hurdles for complex tasks:

  1. Efficiency: Pure systems often suffered from combinatorial explosions during search.
  2. Intelligibility: Fragmenting knowledge into tiny, disconnected rules made it difficult for human experts to understand or debug the system's logic.
  3. Flexibility: A fixed interpreter prevented the system from adapting its strategy to different sub-tasks.

Lenat and McDermott observed that successful systems like DENDRAL (chemistry) and MYCIN (medicine) were effectively "cheating" by incorporating domain-specific structures that violated the pure PS model. They argued that these "cheats" were not flaws, but necessary architectural evolutions.

Proposed Architectural Relaxations

The authors proposed systematically relaxing the constraints on the three core components of production systems:

1. Data Memories (DM)

Instead of a single, flat list of temporary facts, the authors suggested multiple data memories. This includes the use of highly structured, permanent domain knowledge. By allowing data to be stored in complex, interrelated networks, the system’s internal state becomes "statically intelligible." A human observer can look at the data memory and understand the "theory" the system is currently testing, rather than deciphering a pile of disconnected assertions.

2. Production Memories (PM)

In a pure system, rules are a giant, unordered bucket. Lenat and McDermott argued for grouping and prioritizing rules. Furthermore, they suggested that rules should be treated as data—allowing the system to inspect, modify, or even generate new rules. This is essential for systems like AM (Automated Mathematician), which required the ability to discover new mathematical concepts by refining its own heuristics.

3. Interpreters

Perhaps the most radical proposal was making the interpreter "dynamically modifiable." Rather than a hard-coded loop, the interpreter should be a data structure that the rules themselves can influence. This introduces "meta-knowledge"—rules about how to use rules. For instance, if a system identifies it is in a time-sensitive situation, it could modify its interpreter to switch from an exhaustive search to a faster, heuristic-driven approach.

Key Insights: Grain Size and Meta-Knowledge

A central theme of the paper is the "grain size" of knowledge. The authors argue for a "coarse grain size," where a single rule captures a complete "unit of advice" meaningful to a human expert. If a piece of medical wisdom is split into twelve tiny, intercoupled rules to satisfy a "pure" architecture, it becomes impossible to maintain.

Additionally, the shift toward meta-knowledge allows for self-improvement. For an AI to truly learn or discover, it must be able to treat its own strategies as objects of analysis. By relaxing the interpreter's rigidity, the authors paved the way for systems that could reason about their own reasoning processes.

Impact and Limitations

The qualitative shift from "General AI" to "Knowledge-Based Systems" was largely justified by the arguments presented in this work. The benefits of this transition include:

  • Increased Efficiency: Domain-specific wisdom allows systems to prune search trees effectively.
  • Human-Centric Design: Coarser rule grains make systems easier for doctors, chemists, and engineers to program and audit.
  • Discovery: Self-modifying architectures enable automated discovery in complex fields like mathematics.

However, the authors acknowledged that "breaking the rules" involves trade-offs. Moving away from generality means a system tailored for organic chemistry cannot be easily repurposed for another task, such as chess. Furthermore, managing multiple data memories and self-modifying code increases the architectural complexity and the risk of the "retrieval problem," where the system may struggle to find relevant information buried in complex structures.

Conclusion

Lenat and McDermott’s analysis marked a turning point in AI history. By advocating for "less than general" architectures, they provided the theoretical framework for the expert systems that would dominate the 1980s. Their work suggests that while pure architectures are valuable for studying the fundamentals of human cognition, the path to high-performance artificial intelligence lies in embracing the structured, messy, and specialized nature of human expertise.

Summary updated

BibTeX
@inproceedings{lenat1977less,
  author = {Douglas B. Lenat and John P. McDermott},
  title = {{Less Than General Production System Architectures}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure