Experiment #134 · Thought experiment

The Frame Problem

How to update beliefs about what hasn't changed

John McCarthy and Patrick Hayes · 1969 · AI, philosophy of mind

First published: J. McCarthy & P. J. Hayes, "Some Philosophical Problems from the Standpoint of Artificial Intelligence", in *Machine Intelligence* 4 (1969): 463–502.

A robot moves a battery into another room. How does it know the colour of its body, the position of the door, the time of day, and a million other things haven't changed?

In a formal AI representation of action, an agent needs rules not just for what changes when it acts, but for what *doesn't*. Naively encoding "moving a battery doesn't change the room's colour, doesn't change the time, doesn't change..." requires an unmanageable explosion of frame axioms. Dennett popularised the philosophical version: how does any intelligent system efficiently determine which of a vast number of possible consequences of an action are relevant? The problem launched non-monotonic logic, situation calculus, and massive debate over cognitive architecture. Modern machine learning sidesteps the formal problem while inheriting deeper versions of it.

Formulation

Agent performs action A. Question: which facts about the world change as a consequence, and which don't? Formal approaches: explicit frame axioms (intractable), default rules (non-monotonic logic), situation calculus, event calculus. Philosophical version: how does any cognitive system efficiently isolate relevant from irrelevant consequences?

Compare this experiment

Dimensions Engaged

◎

Observer

Observer · Knowledge Extent: how does an agent maintain a working model of the world across time-extended action?

⧉

Information

A challenge to fully-explicit informational models of cognition.

Responses — How Schools Engage

Affirms / takes the bait 3

Naturalism

Modern naturalist cognitive science takes the frame problem as a foundational constraint: cognitive systems must be relevance-sensitive in ways no purely-logical formalisation captures.

Pragmatism

A canonical illustration of the situated, embedded character of cognition. Disembodied formal symbol-manipulation cannot solve the frame problem; intelligence is constitutively contextual.

Phenomenology

The frame problem is what disembodied formal AI sees from outside what embodied agents handle natively. Heideggerian phenomenology was making the point decades before AI rediscovered it.

Reframes the question 2

Structuralism

Treats the problem as identifying the right level of structural abstraction: relevance is a structural property of representations, and the right structure dissolves the formal explosion.

Simulation Theory

A practical constraint on any simulation that purports to model intelligent agents; modern simulators exploit massive parallelism and statistical shortcuts to handle it.