Abstract
In this influential work, Robert Kowalski examines the development of logic programming (LP) as a unifying framework for artificial intelligence. He argues that by extending standard Horn clauses with features such as negation as failure, abduction, and metareasoning, LP can effectively reconcile the historical tension between declarative and procedural knowledge representation.
Introduction: The Declarative-Procedural Divide
For decades, AI research was characterized by a fundamental split. "Declarativists" advocated for representing knowledge as a collection of logical facts, while "proceduralists" viewed knowledge as a set of instructions for problem-solving. Kowalski’s work addresses the limitations of early logic programming, which, while elegant, often lacked the flexibility required for complex AI tasks like non-monotonic reasoning and multi-agent systems.
The Core Paradigm: Logic as Procedure
The foundational "aha moment" of logic programming is the procedural interpretation of logical implications. A rule such as $A \leftarrow B$ is not just a statement of truth; it is an instruction: "To solve problem $A$, solve sub-problems $B$." This interpretation allows logic to serve as a high-level programming language.
As shown in Figure 1, a query like parent(X, jack) can be automatically decomposed into sub-queries for mother and father, eventually grounding the search in specific facts like mary or john. This backward-reasoning strategy provides the efficiency necessary for practical computation.
Extending the Framework
To make logic programming a "complete" tool for AI, Kowalski synthesizes several key extensions:
Normal Logic Programming and Negation as Failure
Standard logic struggles with the "Frame Problem"—the difficulty of representing what remains unchanged after an action. Normal logic programming introduces "negation by failure" (NF), allowing a system to assume a statement is false if it cannot be proven true. This enables efficient non-monotonic reasoning, where the system can maintain defaults that are only overridden by explicit evidence.
Abductive and Metalogic Programming
Abductive Logic Programming (ALP) allows rules to generate hypotheses or explanations for observed facts, rather than just deriving consequences. Metalogic programming further extends this by treating programs themselves as data. By using predicates like demo(Theory, Goal), an AI can reason about the beliefs of other agents, representing them as distinct logical theories.
Constraint Logic Programming (CLP)
CLP improves efficiency by partitioning rules into a general logical component and a specialized "constraint" component. This allows domain-specific solvers (e.g., for arithmetic or linear programming) to handle mathematical constraints more efficiently than general-purpose logical inference.
Key Insights and Applications
Kowalski highlights that logic programs are often interpreted as "if-and-only-if" definitions. This duality supports both backward reasoning (finding causes) and forward reasoning (predicting effects). Furthermore, the structured nature of LP makes it uniquely suited for representing legal reasoning, where laws are essentially sets of definitions and conditions.
Conclusion
While challenges remain—particularly in handling disjunctive logic and full first-order logic—Kowalski’s roadmap suggests that the future of AI lies in the integration of these extensions into a cohesive framework. By combining the rigor of logic with the efficiency of procedural execution, logic programming remains a cornerstone of intelligent system design.
Summary updated