NOTE-351: Intelligence without representation

Contribution

The paper gives an engineering argument, with robots that run, for building intelligence out of parallel competences instead of around a central model of the world. Its two new things are a decomposition and a methodology. The decomposition slices a system “in the orthogonal direction”, into layers that each produce an activity, instead of into perception, a central reasoner and action. The methodology requires a complete, real-world-tested system at every step, with one new layer added and debugged at a time. The architecture itself (subsumption) was published in 1986 and is summarised here, not introduced.

Key insight

If every part of a system connects sensing to action by itself, the system needs no shared description of the world. Each layer reads what it needs straight off its sensors, “projections of a representation into a simple subspace, if you like”, and acts. A newer layer steers an older one only by overriding its signals for a while. So nothing in the system holds a world model, a goal list or a controller, and yet the robot avoids people, wanders and heads for distant places. The unity an onlooker sees is, on Brooks’s account, the onlooker’s: the robot “is a collection of competing behaviors”.

Assumptions

• The goal is engineering, not explanation. “I have no particular interest in demonstrating how human beings work … I have no particular interest in the philosophical implications of Creatures.” The requirements are timely response, graceful degradation, multiple goals that can be switched, and “some purpose in being” (§4).
• Evolutionary argument. Evolution spent most of its time on “the ability to move around in a dynamic environment”. Problem solving, language and reason are “pretty simple once the essence of being and reacting are available” (§2). This is argued from the timeline, not shown.
• Insect-level first. The working target is “simple insect level intelligence within two years” (§8), not human-level.
• Careful engineering of interactions. “We are not claiming that chaos is a necessary ingredient”; interactions are to be carefully engineered (§5.1).

Key results

• The critique of abstraction (§3). AI “succeeds” by defining away the unsolved parts. The researcher does the abstraction, producing atoms like PERSON, CHAIR and BANANAS, “leaving little for the AI programs to do but search”. The human-supplied “Merkwelt” (after von Uexküll) is both the wrong one for robots with other sensors and possibly not the one humans really use. MYCIN, told of a ruptured aorta, looks for a bacterial cause.
• Decomposition by function vs. by activity (§4). Functional decomposition needs “a long chain of modules to connect perception to action”, and none can be tested until all are built, so interfaces are chosen by the module’s own researchers and “subject to intellectual abuse”. Activity decomposition gives “an incremental path from very simple systems to complex autonomous intelligent systems”.
• Who has the representations? (§5). “Not by design, but rather by observation”, the layered robots have no central representation.
  • Low layers react quickly because they maintain no representations.
  • Changes in the world disable some layers, not all.
  • Each layer has “its own implicit purpose (or goal if you insist)”, checked continuously against the world, which is “its own model”.
  • “There need be no explicit representation of goals that some central (or distributed) process selects from.”
  • §5.1: “Just as there is no central representation there is not even a central system.” Coherence emerges “in the eye of an observer”, and “Minsky [10] gives a similar account of how human behavior is generated.”
  • Even locally, “we never use tokens which have any semantics that can be attached to them”. A number passed between processes can be interpreted only “by looking at the state of both the first and second processes”. Brooks grants that “an extremist might say that we really do have representations, but that they are just implicit”, and declines the word.
  • Following Simon’s ant and Agre & Chapman, “much of even human level activity is similarly a reflection of the world through very simple mechanisms” (a hypothesis).
• Methodological maxims (§6.1). Test in the real world, never a simplified one, since a simplified world infects the interfaces. Debug each layer extensively before adding the next, so that only the new layer can be varied.
• The instantiation (§6.2). Four robots of three designs: Allen with an off-board Lisp machine, Tom and Jerry with single PALs, and Herbert with a 24-node CMOS parallel processor. Layers are fixed networks of finite-state machines, with “a handful of states, one or two internal registers, one or two internal timers”, passing 1-bit or 24-bit messages. “All finite state machines are equal, yet at the same time they are prisoners of their fixed topology connections.” The three layers on Allen are:
  1. avoid: twelve sonars, collide, feelforce, runaway, turn, forward;
  2. wander: a random heading every ten seconds or so, summed with repulsion;
  3. explore: whenlook, stereo, pathplan and integrate, steering around the obstacles the lower layer avoids.
• What it is not (§7). It is not connectionism, whose nodes are uniform and which hopes for distributed representations: “we believe representations are not necessary and appear only in the eye or mind of the observer”. It is not neural networks, and “no biological significance” is claimed. It is not production rules, since there are no variables and no matching. It is not a blackboard, since connections are hard-wired. It is not Heideggerian, since it was “based purely on engineering considerations”.
• Limits to growth (§8). Three layers ran on a robot and six in simulation. A fourteen-layer soda-can robot is planned, with behaviours that “index off of the state of the world” instead of being centrally coordinated. Learning works only “as an isolated subsystem”, which is “the very position we lambasted most AI workers for earlier in this paper”.

Claims

id	claim	strength	support
C1	At simple levels of intelligence, explicit world models get in the way; the world is its own best model	moderate for the robots built; untested beyond them	three-layer robot run “for well over a year”; described, no measurements or comparisons
C2	Representation is the wrong unit of abstraction for the bulk of intelligent systems	hypothesis, labelled (H)	the robots, plus the evolutionary timeline argument
C3	Decomposing by activity gives an incremental path to complex autonomous intelligence	weak–moderate	argued against functional decomposition; demonstrated to three layers
C4	The Creature has no central representation, no central control and no explicit goals; its coherence is imputed by an observer	strong for the architecture (true by construction: no global data, no global control); assertion for the interpretive claim about coherence	§§5.1, 6.2
C5	Much of human-level activity reflects the world through simple mechanisms, without detailed representations	hypothesis	Simon’s ant, Agre & Chapman’s unpublished memo; Minsky cited as “similar”
C6	As of mid-1987 these are “the most reactive real-time mobile robots in existence”	the author’s report	no comparison given
C7	Testing in simplified worlds produces systems that must be rebuilt for the real one	moderate as experience; not shown	argued from how interfaces absorb a simplified world’s properties

Concepts

• Creature — a “completely autonomous mobile agent” that coexists with humans and is seen by them “as intelligent beings in their own right”.
• Layer / activity — a behaviour-producing subsystem that connects sensing to action by itself and “must decide when to act for [itself]”, not a subroutine.
• Subsumption architecture — layers of fixed-topology finite-state machine networks, combined by suppression (a new wire replaces an input for a set time) and inhibition (a new wire blocks an output for a set time).
• Merkwelt — the perceptual world of an animal or robot species (after von Uexküll), which AI programs are handed ready-made by their designers.
• The world as its own model — reading the state of the world afresh through sensors instead of maintaining an internal copy.

Connections

• Minsky (LIT-429). Cited directly, and for the point that matters here: “Minsky [10] gives a similar account of how human behavior is generated”, meaning coherent behaviour from competing parts with no central locus. The two differ in what the parts are. Minsky’s agents include memory and representational machinery (K-lines, frames, nemes; LIT-402, LIT-444). Brooks’s layers have “no variables” and pass uninterpreted numbers. Brooks is the more austere decomposition.
• Selfridge and the homunculi. Brooks does not cite Selfridge’s Pandemonium or homuncular functionalism. His competing behaviours with no arbiter sit in the same family, but the lineage is not one he claims.
• Levin (LIT-439). Levin cites a 1986 paper from the same programme (Cudhea & Brooks, “Coordinating multiple goals for a mobile robot”) as evidence that “modular designs with sub-goals” help. He then complains that robots are built from “reliable but very dumb parts”. Brooks’s layers are dumb parts by design, and each has its own “implicit purpose”. The two agree that purpose can be distributed over parts. They disagree on whether the parts must be competent agents.
• IIT 3.0 (LIT-445). Not connected by either author. The bearing is through IIT’s rule that elements outside a candidate set are only background conditions. Brooks’s layers “interface directly to the world through perception and action, rather than interface to each other particularly much”. Coordination that runs through the world would not integrate them on IIT. Whether a subsumption network is one complex, several, or none would depend on its internal wiring, which IIT treats as decisive and Brooks treats as an engineering detail.
• Schwitzgebel (LIT-159, NOTE-131). Not connected by either author. Brooks’s observer-imputed coherence is the deflationary reading of what Schwitzgebel’s telescopic view treats as real. The “Brooks” NOTE-131 lists in the group-mind tradition is a different author, very likely D.H.M. Brooks on group minds (unverified).

Bearing on the record

For the Minsky–Schwitzgebel bridge this is the decomposition side, made concrete. It shows that a whole can pursue several goals, switch between them and degrade gracefully, with no part holding a model, a goal list or control. That is an existence proof at insect scale for Minsky’s thesis that mind-like behaviour needs no mind-like part.

Against NOTE-131’s numbering it is neutral on C3, C4 and C5. It makes no claim about experience in parts or wholes, and its parts are not minded at all. Two of its results are still useful to a THEORY on the seam:

• Goal-directedness without self-representation. Schwitzgebel counts goal-directed responsiveness, self-monitoring and self-representation among the materialist marks the United States has. Brooks’s Creatures have the first with “no explicit representation of goals” and none of the others. So the marks come apart, and a criterion for consciousness that counts them separately will treat a subsumption robot as a partial case.
• Observer-imputed unity. Brooks says the unity of the Creature exists “in the eye of an observer”. A THEORY that treats a composite’s unity as real, as Schwitzgebel’s “entityhood-enough” does, should acknowledge this deflationary reading of the same kind of architecture.

Boundary. The methodological maxims are an instruction for AI and robotics practice, so the LIT is flagged anthology-candidate. The anthology’s agents-and-environments topic could hold it. They are not an instruction for machine-learning practice as such: the paper predates learning in these systems, and its learning module is “isolated”.

Limitations

• Thin evidence. Three layers on one physical robot and six in simulation. Performance is described, never measured or compared, and the most ambitious system (fourteen layers) is a plan.
• Scaling is open by the author’s own account. “How many layers … before the interactions between layers become too complex”, “how complex can the behaviors be” and whether learning is possible are posed in §8 and not answered.
• The no-representation claim depends on a narrow definition. Representations are refused the name because they lack variables, rule matching and choice. Brooks concedes that the system’s numbers and topology could be mapped to a representation. So the claim is about explicit, central, symbolic representation, which is what the title overstates.
• The evolutionary argument is a timeline. That evolution spent longer on mobility than on reasoning does not show that reasoning is “pretty simple” once mobility is solved.
• Unread sources. The 1986 architecture paper and Agre & Chapman’s memo, on which §§5–6 lean, were not read here.

Open questions

• Does activity decomposition scale past a handful of layers without an arbiter? The later behaviour-based robotics literature would answer it. None of it is held in the record.
• Is “coherence in the eye of an observer” compatible with the system being a real agent in Levin’s sense (LIT-439), or with its being a real subject in Schwitzgebel’s? The paper gives the deflationary answer without arguing against the others.
• Under IIT 3.0, what are the complexes of a subsumption network? The answer would test whether the architecture that most closely realises the society of mind is, on that theory, one subject, several, or none.