Rod Brooks:
Intelligence without Representation or Reasoning

• 3 Questions for KRR:
  1. Do we need to represent at all?

  2. If so, what should be represented?

  3. And how?
     • E.g., logic? no logic?

  ---

• Highlights of Brooks's theory:
  1. Traditional AI:
     • top-down approach: intelligence via thought & reason

     • agent senses world;
       builds model of world;
       • with heavy computational load
       • (and only a partial model: Smith 1985)
       plans actions using the model
       • & ignoring the world!
       acts in the world (cf. Smith 1985)

     ---

  2. Bottom-up approach to AI:
     • instead of decomposing "human-level intelligence" into parts (vision, language, etc.) and interfaces...

     • ... start with simpler systems that are "complete"
       • i.e., that can operate in real world with appropriate sensory & acting abilities
       • e.g., robotic insects

     • i.e., follow evolutionary path:
       • build mobility, vision, survival tasks as foundation for "true intelligence"

       • "these are the hard tasks"...
         • most of intelligence "is routine activity in a...benign but...dynamic world"

       • The hard tasks "constrain" solutions to the routine problems
         • viz., the classic AI problem-solving & planning problems:
         • e.g., games, math, logic
           • Cf. basic trade-off paradox of AI:
             Tasks that are
             • hard for people to do
             • but easy to teach (games, math, logic, etc.)
             • are easy to develop computational theories of,
             • but tasks that are easy for people to do
             • but hard to teach (e.g., vision, mobility, survival)
             • are hard to develop computational theories of.

       • "Thought & consciousness don't need to be programmed in; they will emerge"

     ---

  3. "The world is its own best model"
     • use world as its own model by continually monitoring & sensing the world

     • "explicit representations & models of the world...get in the way"

     • "internal world models...are...impossible to obtain [&] not...necessary" for action

     ---

  4. "Representation is the wrong unit of abstraction in building the bulkiest parts of intelligent systems"
     • so, possibly representation is OK for some things:
       • language
       • problem solving
       • reasoning
       • memory

     • see below

     ---

  5. "Representations...appear only in the eye or mind of the observer"
     • = Dennett's "intentional stance"

     • possibly, representations are ways for a researcher to describe what a robot does & how it works
       • "observer" can...talk about an agent's beliefs & goals, even though the agent need not manipulate symbolic data structures"

     • COG

     • But cf. Brian Cantwell Smith's Knowledge Representation Hypothesis

     ---

  6. Brooks's "key aspects":
     • "robots are situated in the world"
       • "experience the world directly"
         
       • physical grounding gives meaning to system's processing

     • "are observed to be intelligent"

     • "intelligence...emerges from system's interaction with the world"
       
       • cf. Simon's ant

     ---

  7. Real question:
     • What is the role of thought in action?
       • as opposed to behavior

     • memory, reasoning, advice taking (learning by being told), etc., all require concepts & therefore representations

     • at what level do representations "emerge" or become necessary?
       • Brooks: "partial models of the world" that are "relevant" to the current task (e.g., maps)

       • Agre & Chapman: there are "deictic" representations for specific purposes only.
         • e.g., not: Bee-21
         • but: the-bee-that-is-chasing-me-now
         • which may/may not be the same bee chasing me before
           • deictic representations are intensional??
           • (Cf. memory not as storage but as re-firing of neurons)

     ---

  8. What about reasoning?
     • Brooks: not needed (but he doesn't give much of an argument)

     • Possibly:
       • There are no permanent representations
       • Therefore, there is nothing to reason about or with?

     • Probably:
       • Intelligence may be less a function of reasoning than of interaction of system components with themselves and with the world
         (citing Steels)

     • But: Gentner 2003: We're smart because of our analogical reasoning abilities operating on KRR systems.

  ---

• References:
  1. Dennett, Daniel C. (1971), "Intentional Systems", Journal of Philosophy 68(4): 87-106; reprinted in Daniel C. Dennett, Brainstorms (Montgomery, VT: Bradford Books): 3-22.
     • For some interesting followups to Dennett's paper, see:
       Miller, Christopher A. (guest ed.) (2004), "Human-Computer Etiquette: Managing Expectations with Intentional Agents", Communications of the ACM 47(4) (April): 31-34.

  2. Brooks, Rodney A. (1991), "Intelligence without Representation", Artificial Intelligence 47: 139-159.

  3. Brooks, Rodney A. (1991), "Intelligence without Reason", IJCAI-91 (San Mateo, CA: Morgan Kaufmann): 569-595.

  4. Kirsh, David (1991), "Today the Earwig, Tomorrow Man?", Artificial Intelligence 47: 161-184.

  5. Brooks, Rodney A. (1996), "From Earwigs to Humans", Proceedings IIAS: The Third Brain and Mind International Symposium Concept Formation, Thinking and Their Development, Kyoto, Japan: pp. 59-66.

  6. Gentner, Dedre (2003), "Why We're So Smart", in Dedre Gentner and Susan Goldin-Meadow (eds.), Language in Mind: Advances in the Study of Language and Thought (Cambridge, MA: MIT Press): 195-235.