Learning: The Construction of A Posteriori Knowledge Structures

Paul D. ScottAAAI 1983

Author: Paul D. Scott (1983)

Abstract

In the early 1980s, the field of artificial intelligence lacked a rigorous, unified definition of learning. This paper critiques the prevailing functionalist view—which equates learning with performance improvement—and proposes a structural alternative. By defining learning as the construction of synthetic a posteriori knowledge structures from experience, the author provides a theoretical framework that bridges the gap between machine learning and knowledge representation.

Introduction

During the artificial intelligence "renaissance" of the early 1980s, researchers struggled to distinguish between simple system optimization and true machine learning. The dominant perspective, championed by figures like Herbert Simon, was functional: learning was defined as any change in a system that allowed it to perform a task more effectively or efficiently upon repetition. In "Learning: The Construction of A Posteriori Knowledge Structures," Paul D. Scott argues that this definition is both philosophically and practically inadequate for the development of intelligent systems. Instead, he posits that learning is fundamentally about the organization of experience into internal representations.

The Problem with Functional Definitions

Scott identifies two primary conceptual failures in the performance-based definition of learning:

  1. The Razor Blade Problem: If a dull blade in a manual razor is replaced with a sharp one, the razor performs its task better. Under a purely functional definition, the razor has "learned." This contradicts the intuitive understanding of learning as an internal cognitive or computational process.
  2. The "Useless" Fact Problem: A system may acquire information that is true and derived from experience—such as a trivial biographical detail about an author—without that information ever contributing to a specific performance metric. To claim the system has not "learned" simply because the knowledge is not currently being utilized ignores the reality of latent knowledge acquisition.

To resolve these issues, Scott suggests shifting the focus from how well a system performs to what kind of knowledge it constructs to represent its environment.

A Structural Framework: Knowledge Builders and Users

Scott introduces a conceptual architecture to clarify the learning process, dividing an intelligent system into two distinct components:

  • Knowledge Users (KU): These are the subsystems that utilize existing information to execute tasks and interact with the environment.
  • Knowledge Builders (KB): These are the subsystems responsible for monitoring experience and constructing or maintaining the knowledge structures used by the KU.

Within this framework, learning is defined as the acquisition of synthetic a posteriori knowledge. This refers to information derived from the environment that cannot be deduced through pure logic alone. By focusing on the "Knowledge Builder," Scott argues that learning is essentially the process of moving from specific, messy environmental data to organized, generalizable structures.

Key Insights and Implications

One of the paper’s central tenets is that learning and knowledge representation are inseparable. A system’s ability to learn is strictly bounded by its representational language. Consequently, AI research cannot treat learning algorithms in isolation; they must be designed in tandem with the structures they are intended to build.

Capturing the Unformalizable

The true value of machine learning lies in its ability to capture "unformalizable" expertise. Scott notes that human experts, such as chess grandmasters, often possess knowledge they cannot explicitly articulate as rules. Learning allows a machine to bypass the "knowledge acquisition bottleneck" by extracting this messy, intuitive expertise directly from experience rather than relying on human programmers to formalize it.

The Copying Advantage

Scott highlights a significant divergence between biological and artificial learning: the ability to replicate. While every human must undergo a tedious, individual learning process to acquire a skill, a machine representation—once constructed by a "Knowledge Builder"—can be instantly copied to an infinite number of other machines. This makes the initial "construction of knowledge" a high-leverage activity for AI development.

Reinterpreting Classic AI

To demonstrate the utility of his framework, Scott re-evaluates Arthur Samuel’s famous 1963 Checkers program. Rather than viewing it merely as a function optimizer adjusting coefficients, Scott interprets it as a system building a model of an opponent’s behavior. The program operated on the a priori assumption of rational play (minimax) and used experience to fill in the a posteriori details of the specific game state evaluation.

Conclusion

Scott concludes by calling for a new taxonomy of machine learning. Rather than classifying programs by their underlying algorithms (e.g., neural networks vs. symbolic logic), he argues they should be classified by the nature of the knowledge they acquire. This shift necessitates a move away from purely deductive logic toward a more robust exploration of inductive logic—the rules and heuristics that allow a system to transform specific experiences into enduring knowledge structures. The paper remains a foundational text for understanding learning not as a mere performance gain, but as the architectural evolution of an agent's internal world.

Summary updated

BibTeX
@inproceedings{scott1983learning,
  author = {Paul D. Scott},
  title = {{Learning: The Construction of A Posteriori Knowledge Structures}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}

Loading related content...

Figure