Meta-Level Knowledge in Expert Systems: Davis and Lenat's Formalisation at Stanford
In 1977, Randall Davis and Douglas Lenat at Stanford University published research formalising meta-level knowledge, a system's explicit representations of its own knowledge and reasoning strategies, and demonstrated its application in the MYCIN and AM expert systems to improve inference control and self-directed learning.
Background
By the mid-1970s, expert systems were beginning to show real promise. MYCIN, built at Stanford University to diagnose bacterial infections and recommend antibiotics, could match the performance of specialists in its narrow domain. AM, another Stanford project, could rediscover basic mathematical concepts by searching through a space of possible ideas using heuristics, rules of thumb about what directions looked promising.
Both systems worked. But they worked in a way that made them hard to inspect or improve. The knowledge they used to reason, and the knowledge they used to control that reasoning, sat tangled together. If a system made a poor inference, or spent time chasing unproductive lines of thought, there was no clean way to ask why. The strategies guiding the system were implicit, buried in code rather than represented as something the system itself could examine or adjust.
This created a ceiling. You could add more facts to a knowledge base, but you could not easily teach a system to reason better, because the reasoning strategies were not objects the system could see or work with.
What happened
Randall Davis and Douglas Lenat, both working within Stanford University’s Heuristic Programming Project, published research in 1977 that addressed this directly. Their paper, presented at the International Joint Conference on Artificial Intelligence, introduced a formal treatment of what they called meta-level knowledge: a system’s explicit representations of its own knowledge and of the strategies it uses to reason.
The core idea was a separation of levels. Object-level knowledge covered facts about the domain, the kind of thing MYCIN knew about bacteria and drugs. Meta-level knowledge covered the system’s knowledge about that knowledge: which rules were reliable, when a line of reasoning was worth pursuing, how to choose between competing inference paths. By making this second layer explicit, a system could reason about its own behaviour in something like the way it reasoned about external facts.
Davis and Lenat demonstrated this in the context of both MYCIN and AM. In MYCIN, meta-level representations gave the system a way to control its own inference process, explaining its reasoning to users and handling uncertainty more transparently. In AM, the heuristics that guided concept discovery were represented as first-class knowledge, things AM could inspect rather than just execute. The paper showed that both uses were instances of the same underlying architectural idea, and gave that idea a precise definition and a name.
The work came out of a 1976 Stanford technical report, HPP-76-7, before appearing in the IJCAI-77 proceedings. This meant the ideas had already circulated within the research community before formal publication, which was common for influential work coming out of the Heuristic Programming Project at the time.
Why it mattered
Davis and Lenat's framework established meta-level knowledge as a distinct architectural concept in AI, separating object-level domain facts from the strategies a system uses to reason about them. This distinction influenced the design of later expert systems and early reflective architectures. The work also contributed to understanding how a system might guide its own learning, a concern that remained central to AI research for decades.
People
Randall Davis, Douglas B. Lenat
Organisations
Stanford University, Stanford Heuristic Programming Project
Sources
- Meta-Level Knowledge: Overview and Applications.ACM Digital Library / IJCAI-77 Proceedings.Primary source
- Meta-Level Knowledge: Overview and Applications (Stanford Heuristic Programming Project Memo HPP-76-7).Stanford University Libraries.institutional
- Stanford Artificial Intelligence Laboratory Records, Heuristic Programming Project.Stanford University Libraries Special Collections.institutional
Cite this page
AI Achievements. (1977). Meta-Level Knowledge in Expert Systems: Davis and Lenat's Formalisation at Stanford. Retrieved 2026-08-22, from https://achievements.ai/milestone/meta-level-knowledge-randall-davis
@misc{achievements_meta_level_knowledge_randall_davis,
title = {Meta-Level Knowledge in Expert Systems: Davis and Lenat's Formalisation at Stanford},
author = {{AI Achievements}},
year = {1977},
url = {https://achievements.ai/milestone/meta-level-knowledge-randall-davis}
}