Abstract
This seminal 1983 work by Michael Genesereth proposes a "meta-level architecture" for artificial intelligence, which separates a program's core actions (base-level) from the logic that decides which action to take (meta-level). By using a declarative language to specify control, researchers can build flexible, incremental systems that unify disparate AI paradigms—such as search, object-oriented programming, and "demons"—into a single, cohesive framework.
Introduction: The Control Problem
In the early development of AI, a primary challenge was the "control problem": determining how to guide a program's search for a solution without hard-coding every decision. Traditional programming often entangles domain logic (the "what") with control strategy (the "how"), resulting in brittle systems that are difficult to modify or integrate.
Genesereth argues for the necessity of "partial specifications" of behavior. Rather than defining a rigid algorithm, programmers should provide "advice" or constraints on program behavior. This allows a system to evolve from a general, slow searcher into a highly efficient, specialized tool through the incremental addition of control knowledge.
The Meta-Level Framework
The paper introduces a dual-agent model to decouple domain actions from control logic.
- The Base-Level Agent: This agent perceives and modifies the external environment (e.g., manipulating blocks in a simulation).
- The Meta-Level Agent: This agent treats the base-level agent as its environment. Its objective is to observe the base-level's state and goals, then determine the "ideal" action for the base-level to perform next.
The core of this architecture is a Declarative Control Language based on predicate calculus. Instead of nested procedural loops, the programmer writes logical axioms. Key components include:
- Vocabulary of Actions: Symbols such as
OPR(operator),IN/OUT(inputs/outputs), andTIMEallow the meta-level to reason about tasks. - The Scheduler: A simple interpreter (originally LISP-based) that executes a loop: it identifies a binding for a variable
zthat satisfies the propositionRECOMMENDED(z), executes that task, and repeats. - Axiomatization of Paradigms: Different AI styles can be "wrapped" into this logic. For example, a "demon"—code that triggers on a specific condition—is represented as an axiom stating that if a condition is met, a specific task is
APPLICABLE.
Key Insights and Unified Paradigms
The architecture offers several conceptual advantages:
- Control as Data: By representing control strategies as logical axioms, the program's "intelligence" regarding problem-solving becomes inspectable and modifiable data.
- Search Control via Preference: The introduction of the
PREFERRED(x, k)relation allows for elegant search control. For instance, depth-first search can be implemented by an axiom that prefers tasks deeper in the goal tree. - Incremental Development: Systems can begin with basic functionality and gain efficiency as meta-level axioms (control advice) are added, without requiring a rewrite of the base logic.
- Unified Architecture: The author demonstrates that Object-Oriented Programming and traditional procedural programming are simply different ways of writing meta-level axioms, allowing them to coexist within the same system.
Implementation and Results
While primarily conceptual, the paper cites several successful implementations of this architecture:
- DART: An automated diagnostician for computer hardware faults.
- MINIMA: A program for solving calculus problems.
- NEOMYCIN: A system for infectious disease diagnosis and tutoring.
- MRS: A knowledge representation system designed to facilitate building meta-level programs.
The qualitative result is a "cleaner" approach to AI construction. The architecture makes programs more aesthetic, easier to build, and significantly more modular than monolithic procedural code.
Limitations and Future Work
Despite its advantages, the meta-level approach faces several hurdles:
- The "Meta-Tax": Reasoning about actions takes computational time. If meta-level reasoning is as complex as the problem itself, the overhead can negate efficiency gains.
- The Compilation Challenge: To mitigate overhead, Genesereth suggests "compiling" meta-level axioms into efficient base-level code. However, building a compiler for arbitrary logical specifications remains a difficult research task.
- Scaling: The model currently assumes a single meta-level and base-level agent. Expanding this to handle multiple interacting agents is a key area for future research.
- Universal Rules: There is a continuing need to discover domain-independent control rules that can be applied across diverse AI applications.
Summary updated