Epistemological Adequacy in McCarthy and Hayes's AI Framework

This article explores the concept of epistemological adequacy, a foundational principle introduced by John McCarthy and Patrick Hayes in their 1969 paper on artificial intelligence. It explains how a representation of the world is considered epistemologically adequate if it can practically express the facts available to an observer. The text below breaks down the core definition of this concept, contrasts it with heuristic adequacy, and highlights its role in knowledge representation, logic, and the famous frame problem in AI.

The Core Definition

In 1969, computer scientists John McCarthy and Patrick Hayes published "Some Philosophical Problems from the Standpoint of Artificial Intelligence." In this work, they divided the problem of artificial intelligence into two distinct parts: the epistemological part and the heuristic part.

According to their framework, a representation of the world is epistemologically adequate for a person or machine if it can be used practically to express the facts that the entity actually has about the aspect of the world being examined.

Crucially, an epistemologically adequate representation does not require complete, omniscient knowledge of the universe. Instead, it must accurately reflect the limitations of an agent's knowledge. For example, if a robot does not know the exact location of a key, an epistemologically adequate system allows the robot to represent the fact that the key is "either in the drawer or on the table," without forcing the robot to commit to a specific x-y-z coordinate.

Epistemological vs. Heuristic Adequacy

To fully understand epistemological adequacy, it must be separated from its counterpart: heuristic adequacy. McCarthy and Hayes made this distinction to separate what an AI needs to know from how an AI processes that information to solve problems.

Epistemological Adequacy: Focuses on the representation of facts, causality, time, and the capability to express partial knowledge. It asks: "Can this language express the reality the AI perceives?"
Heuristic Adequacy: Focuses on the computational efficiency and search processes used to solve problems based on the available facts. It asks: "Can the AI use this representation to find a solution in a reasonable amount of time?"

A system might be epistemologically adequate because it contains all the necessary logical statements to describe a chess game, yet be heuristically inadequate because it takes a million years to compute the best next move.

The Role of the Situation Calculus

To achieve epistemological adequacy, McCarthy and Hayes introduced the "situation calculus." This is a logical formalism designed to represent changing worlds. In this framework:

Situations: Represent a snapshot of the world at a specific point in time.
Fluents: Represent conditions or properties that can change from one situation to the next (e.g., whether a light is on or off).
Actions: Represent events that transform one situation into a new situation.

Using first-order logic, the situation calculus provided an epistemologically adequate way for an AI to express cause and effect, the passage of time, and the results of specific actions without needing to know the entire physical state of the universe.

Connection to the Frame Problem

The pursuit of epistemological adequacy directly led McCarthy and Hayes to discover the Frame Problem, one of the most famous challenges in artificial intelligence.

When using logic to represent changes in the world, it is easy to state what changes when an action occurs (e.g., if a robot paints a block red, the block's color changes). However, to remain epistemologically adequate, the representation must also account for everything that does not change (e.g., painting the block does not change its weight, the location of the robot, or the current president of the United States).

Writing down every single non-change (frame axioms) makes a system clumsy and unusable. Therefore, achieving true epistemological adequacy requires finding a way to represent the persistence of objects and states over time naturally and concisely, a challenge that eventually led to the development of non-monotonic logic in AI.