BEINGS: Knowledge as Interacting Experts

Douglas B. LenatIJCAI 1975

TL;DR

This paper introduces "Beings," a novel AI architecture where knowledge is organized as a community of specialized, interacting modules (experts). This approach aims to simulate how human experts collaborate to solve complex problems, demonstrating its effectiveness in automatic programming by synthesizing a concept formation program.

What Problem Does This Paper Address?

The paper tackles the challenge of organizing and utilizing complex knowledge within an AI system, drawing an analogy to how human experts collaborate. Traditional AI systems often struggle with integrating diverse knowledge sources and managing complex problem-solving processes. The authors propose that by modeling AI components as "Beings" – each an expert in a narrow domain with a standardized internal structure – they can achieve more robust and flexible knowledge representation and interaction, mirroring the synergy of a human expert panel.

Methods

The core innovation is the "Being" concept: a modular AI agent with a standardized set of "parts" (representing questions it can answer or actions it can perform). These Beings interact by posing questions to each other, simulating a dialogue. A key constraint is that all Beings share the same "anatomy" – the same set of possible parts – allowing for uniform communication protocols. This uniformity simplifies inter-Being communication and facilitates the assembly of Beings into executable programs. The system, named PUP6, was implemented to test this hypothesis, focusing on automatic programming.

Datasets

Not discussed in the paper. The paper focuses on the architecture and its application to automatic programming, using simulated dialogues and a synthesized concept formation program as its primary "data."

Key Insights and Ideas

  • Knowledge as a Community of Experts: The most significant idea is framing AI knowledge not as a monolithic entity but as a collection of specialized, interacting agents, much like a human expert panel.
  • Standardized "Anatomy" for Beings: The concept of a fixed, universal set of "parts" for each Being is crucial. This standardization enables seamless communication and allows for a uniform mechanism to assemble Beings into functional programs.
  • Dialogue as the Primary Interaction Mechanism: Beings communicate through a structured question-and-answer protocol, mimicking human expert discussions. This dialogue-driven approach allows for emergent problem-solving and knowledge integration.
  • Automatic Programming via Expert Simulation: The paper demonstrates that this "expert panel" approach can be used for automatic programming. By simulating the process of human experts collaborating to write a program, PUP6 can synthesize new code.
  • The "Frame Problem" Analogy: The paper implicitly touches on the frame problem by emphasizing how Beings manage knowledge and avoid unnecessary computations, though not explicitly framed as such.

Key Results

The experimental system, PUP6, successfully synthesized a concept formation program (CF) that was similar to one described by P. Winston. The synthesized CF program was significantly smaller (20 pages) than the PUP6 system itself (200 pages). The dialogue between the user and PUP6 to generate CF involved 100K characters typed by PUP6 and 4K by the user, occupying 300 pages and taking five hours of real time. PUP6 also synthesized a grammatical inference program and a simple property list manipulator with minor modifications.

Limitations and Future Work

The authors acknowledge several limitations:

  • Brittle User Interaction: The system's dialogue flexibility was limited, heavily relying on the user adhering to the specific interaction patterns of the simulated experts. Real users often misunderstood Beings' requests.
  • Minimal Knowledge Filling: Beings were often implemented with only the knowledge necessary for the specific task, leading to difficulties in handling new dialogues or alternative solution paths.
  • Lack of Debugging: The system lacked a dedicated debugging specialist, making it difficult to handle errors in the synthesized programs.
  • Inefficiency: The "open floor" communication mechanism, while flexible, could be inefficient, requiring time proportional to the number of Beings in the system.

Future work suggested includes:

  • Adding debugging specialist Beings.
  • Developing dialogue specialists and sophisticated user psychology experts.
  • Creating Beings to aid untrained users in inserting new domain-specific knowledge.
  • Building a library of varied specialist Beings.
  • Investigating tasks like "research in elementary number theory" to isolate representation problems from natural language complexities.

Summary updated

BibTeX
@inproceedings{lenat1975beings,
  author = {Douglas B. Lenat},
  title = {{BEINGS: Knowledge as Interacting Experts}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure