How To Use What You Know

Carl HewittIJCAI 1975

Introduction: The Logic-Procedural Divide

In the early 1970s, artificial intelligence research was polarized between two philosophies: pure logic (predicate calculus) and procedural programming. Logic-based systems offered formal rigor but suffered from the "Gigo" (Garbage In, Garbage Out) problem—a single contradiction could collapse the entire system. Conversely, procedural systems were often too rigid and struggled with general reasoning.

Carl Hewitt’s 1975 paper, "How To Use What You Know," proposes a synthesis. He argues that knowledge must be "procedurally embedded"—not merely stored as static facts in a database, but integrated as active participants in problem-solving. To achieve this, Hewitt introduces PLASMA (Planner-like System Modeled on Actors), a framework built on the Actor model.

The PLASMA Architecture

The core innovation of PLASMA is the unification of data and procedure into a single primitive: the Actor. In this model, every entity is an Actor capable of receiving messages and determining its own response. This approach moves beyond simple pattern matching to "World Directed Invocation," where a specific context or "micro-world" triggers relevant plans.

PLASMA incorporates several sophisticated mechanisms for knowledge representation:

  • Stereotypes (Frames): Inspired by Marvin Minsky, these are templates for common situations that include default objects and relations, which can be overridden by specific case data.
  • Nested Continuation Control: Instead of the messy backtracking found in earlier languages, PLASMA uses "then:" and "else:" continuations to manage success and failure (the latter handled by specialized "Complaint Departments").
Formal representation of a state transition in PLASMA, showing the relationship between logical assertions and the physical state.

Planning: From Utopia to Reality

Hewitt introduces a goal-oriented planning strategy that compares "Utopia" (the desired goal state) with "Reality" (the current state). By performing Constraint Analysis, the system works backward from the goal to identify necessary intermediate steps.

When a plan fails, the "Complaint Department" does more than just signal an error; it provides diagnostic information explaining why the failure occurred, allowing the system to pivot to a more viable alternative.

Evaluation in the Blocks World

The paper evaluates PLASMA using the "Blocks World," a classic AI environment of virtual blocks on a table. A key benchmark was the "Brown Blocks World Problem," a complex configuration that defeated linear planners of the era by requiring excessive backtracking.

A stack of blocks (A on B on C), representing a typical configuration used to test planning algorithms in the Blocks World.

PLASMA successfully solved these problems by treating facts as situational. Rather than seeking universal truths, the system uses "situational tags," allowing it to reason about contradictions (like a flightless bird) as special cases rather than logical failures.

Conclusion and Future Directions

Hewitt’s work demonstrates that the declarative vs. procedural debate is a false dichotomy. By embedding knowledge within Actors, PLASMA achieves a robustness to contradiction and a flexibility in planning that pure logic systems lacked. While the initial implementation of World Directed Invocation was nascent, the paper laid the groundwork for applying these "Programming Apprentice" tools to large-scale real-world systems, moving AI toward more human-like, context-aware reasoning.

Summary updated

BibTeX
@inproceedings{hewitt1975how,
  author = {Carl Hewitt},
  title = {{How To Use What You Know}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure