Dan Dennett and the Framing Problem

I'm reading Dennett's article, "Cognitive Wheels: The Frame Problem of AI."

The frame problem is, in the context of computer science and the study of artificial intelligence, a bit of a programming riddle. Basically, the big question is - how do you get a robot to filter out unnecessary information reliably and efficiently, in a manner that actually resembles what we call intelligence, when planning out its actions?

The issue is that if you program a robot to do seemingly simple in-world problem-solving, the robot will run into a number of problems when planning its actions. For example: if it wants to get into a room to take something out, it will often have to figure out all sorts of things that seem simple to us, like how hard to push the door, what the likelihood of finding certain items in the room is, whether it will have to execute particular actions or go particular distances to maneuver into and out of the room, and even updating its memory stores (delete from one place, add to another) whenever any item is moved, etc. It will also have to compute what sorts of conclusions (to use Dennett's example, that there is no elephant in the room) to deem irrelevant to the task at hand.
One way to pseudo-solve the problem is to equip the robot with a bunch of "experience" (read: program a bunch of information into it in the form of axioms that it can derive more information from) relevant to the task at hand. A programmer *could* simply program a robot to do the exact same thing given a set of constants and a variable or two. Nonetheless, it still seems like if you try to give a robot a novel problem to solve, it may have to start making some calculations about what kind of information it needs to use, and then it seems like, in ruling out the infinite things it could do but shouldn't, it will never get going.

At least, that's roughly the way Dennett put it.
Ultimately, this is a combination of a storage problem (how much information is enough, and what is the most efficient way to store it?) and an attention problem (given vast stores of available data, what sort of information is actually relevant to the task at hand, and can you pull it out / conclude deductions quickly enough to be useful?).

To quote Minsky:

Even if we formulate relevancy restrictions, logistic systems have a  problem using them. In In any logistic system, all the axioms are necessarily "permissive" - they all help to permit new inferences to be drawn. Each added axiom means more theorems; none can disappear. There simply is no direct way to add information to tell such a system about kinds of conclusions that should not be drawn... If we try to change this by adding axioms about relevancy, we still produce all the unwanted theorems, plus annoying statements about their irrelevancy. (Minsky, 1981, p. 125) 

Sounds like a tough nut to crack.
"What is needed is a system that genuinely ignores most of what it knows," says Dennett in this article. Precisely! That's why I called it partly a problem of attention. Attention helps one focus on what *is* important, and devote as few resources as possible to that which is not (ie, ignore what's not important). That's not to say that giving a label to the thing we need is any solution to the problem, mind. As Dennett points out, we know what we need at the 'phenomenological level', but not at the nuts-and-bolts level.

One interesting thing to note is the important role boredom plays in human reasoning. If we couldn't get bored (as computers don't) we really *would* follow arguments to irrelevant conclusions all the time - if we weren't bored or frustrated, perhaps we'd never be driven to creativity. Hofstadter (1982), I've been told, has a book out on the importance of boredom to memory and creativity.


Back to the framing problem - apparently equipping a system with a set of stereotypes (basic assumptions about the structure of the world and relations of things in it) helps a little bit. Nonetheless, it seems to have its limitations, with very antagonistic environments producing what Dennett calls "insectlike behavior", namely the dumb repetitiveness of a non-intelligent system trying to solve a problem the same way it has tried to solve it before.

If you ask me, I'm with Gibson on this one - the perceptual systems of the human animal (and animals in general) are the place to get clues. Dennett waves off Gibson's 'affordances' as a postulation of "miracle tissue" - attributing those properties to the visual system is still not quite solving the problem of how those properties emerge from their component parts.

Ladies and gents, do I hear trumpets in the distance?! This is where I come in. I'm going to do a little dance with Monsieurs Prinz and Barsalou and come up with something good; maybe even something better!

Paraconsistent Logic

So, this is just going to be a sort of mental note to myself, but I've recently been interested in the following problem:

Can you convince a person, of any given thing, if they show no commitment to logical consistency?

Let me give an example. For this example, let us assume that "if Z, then X" is actually the case in this world; that is, X really follows from Z. Your interlocutor has granted to you that Z is true, but fails to acknowledge X.
You tell a person, "Oh, but that's not right, the truth is actually that X. See, it follows naturally from 'if Z, then X'."
The person replies, "No, it is the case that not-X."
"But," you say, "You have previously acknowledged that not-X is in direct contradiction of this other belief that you have, Z. And you also acknowledge that if Z, then X. I have given you, and you accept, Z. Why won't you accept X? Isn't that inconsistent?"
And they reply, "Oh I know it's inconsistent, but I really don't care about that."

How would you reply to this person? Would you characterize them as irrational, or just arational?
If they do not at all care about consistency, the whole logical apparatus falls apart. If they're okay with inconsistency, it means that " P & not-P " can be the case. Usually this would work to produce a reductio ad absurdum, but if your interlocutor is unperturbed, well, by golly! what was the point of trying to convince them in the first place?

As it turns out, there are a couple of people (most notably in Australia) working on something called Paraconsistent Logic, which is logic that can deal with contradictions.
I wonder, though, for this logic -
a) Does *everything* follow from P & not-P? Or how do you go about discriminating about what should and should not follow?
b) Would it still work for convincing someone who doesn't care about consistency?

This is an area to look into.

Lewis and Radical Interpretation

Briefly: Lewis wants to know what, given P (where P is the totality of facts about Karl, his subject, as a physical system), would enable us to solve for the rest of Karl's psychological makeup: his beliefs, desires, the meaning of his statements, and anything else that could be said following these things.


Put another way, Lewis wants to know how beliefs, desires and meanings are normally related to one another, and to sensory input and behavioral output. He believes it is possible to construct a theory that adequately accounts for all of these, given P.


"It must amount to nothing more than a mass of platitudes of common sense," he says of the theory. Why must this be the case? Because Lewis is convinced that the fundamental principles of our common-sense theory of persons (that's a mouthful) already implicitly define belief, desire and meaning.
Perhaps it is just a case of weeding these implicit definitions out and making them explicit?


He does believe that this is the case. Nonetheless, the first formulation of the method he has in this paper, even by his own admission, places too much emphasis on language as a vehicle for belief, and not enough on language as a social practice or behavior as a manifestation of belief.  I agree heartily on this latter account. He goes on and gives two more methods of discovery that depend less on these sorts of things, the second of which I believe to be most like the way we naturally go about trying to determine belief.


Agustín Rayo has asked us: "Lewis thinks there is a tight connection between folk psychology and rationality, and accuses his critics of not paying enough attention to this connection. Is he right?"


Perhaps he is partly right. Perhaps the kind of theory he wants is one that science would not necessarily give us, but that would be a systematization of the way we ordinarily go about ascribing belief. Nonetheless, I'm not quite sure why he needs such a systematization in the first place. Is it that the concepts of belief, desire and meaning are so important that, even if science were to break them up into component parts that did not remotely resemble the conception of them in the human mind, we ought to stubbornly find a way to keep using them? What justification is there for this view of his?
And a further question: why challenge people to systematize a method we all use for purely socially beneficial purposes? Why does there have to be a science to belief ascription?






Sept. 18th enmendation:
Lewis' paper (or section) called "Reduction of Mind" is quite brilliant in its discussion of the failures of contemporary philosophy to address narrow content. I also like most of what he has to say about beliefs. If any of you are interested, it is worth a read.