AGRE@SPAR-20.ARPA (Philip E. Agre) (12/14/88)
[Nick -- Following our conversation back in October, I have dug the enclosed bit of text out of an unpublished paper of mine (an AI Lab Working Paper as it happens). If you think it would be appropriate for AIList feel free to include it. I tend to write in two-page chunks, so I haven't really got anything much shorter. Hope it's going well with you. Ciao. Phil] In the old days, philosophers accused one another of associating with a sneaky individual called a *homunculus*. From Latin, roughly ``little person.'' For example, one philosopher's account of perception might involve the mental construction of an entity that `resembled' the thing-perceived. Another philosopher would object that this entity did nothing to explain perception since it required another person, the homunculus, to look at it and determine its identity and properties. Philosophers have been arguing about this issue for centuries. Computational ideas have a natural appeal to philosophers who care about such things because they let one envision ways of `discharging' the homunculus by, for example, decomposing it into a hierarchy of ever-dumber subsystems. I think, though, that the tangled ruts of argument about homunculi distract attention from a more subtle and telling issue. If the homunculus repeats in miniature certain acts of its host, where does it perform these acts? The little person lives in a little world, the host's surroundings reconstructed in his or her head. This little world, I decided, deserves a Latin word of its own. So I talked to my medievalist friend, who suggested *orbiculus*. One way to say ``world'' is *orbis terrarum*, roughly ``earthly sphere.'' So the orbiculus is one's world copied into one's head. Where can we find orbiculi in AI? All over. A `world model' is precisely an orbiculus; it's a model of the world inside your head. Or consider the slogan of vision as `inverse optics': visual processing takes a retinal image and reconstructs the world that produced it. Of course that's a metaphor. What's constructed is a *representation* of that world. But the slogan would have us judge that representation by its completeness, by the extent to which it is a thorough re-presentation of the corresponding hunk of world. You'll also find an orbiculus almost anywhere you see an AI person talk about `reasoning about X'. This X might be solid objects, time-extended processes, problem-solving situations, communicative interactions, or any of a hundred other things. The phrase `reasoning' about X suggests a purely internal cognitive process, as opposed to more active phrases like `using' or `acting upon' or `working with' or `participating in' or `manipulating' X. Research into `reasoning about X' almost invariably involves all-out representations of X. These representations will be judged according to whether they encode all the salient details of X in such a way that they can be efficiently recovered and manipulated by computational processes. In fact, discussions of these `reasoning' processes often employ metaphors of actual, literal manipulations of X or its components. In practice, the algorithms performing these abstract manipulations tend to require a choice between extremely restrictive assumptions and gross computational intractability. If you favor slogans like `using X' over slogans like `reasoning about X', someone will ask you, But we can reason about things that aren't right in front of us, can't we? Mentalism offers seductively simple answers to many questions, and this is one of them. According to mentalism and its orbicular slogans, reasoning about a derailleur proceeds in the same way regardless of whether the derailleur is in front of you or in the next room. Either way, you build and consult an orbicular derailleur-model. If having the derailleur there helps you, it is only by helping you build your model. This is, of course, contrary to common experience. The first several times you try to reason about a derailleur it has to be sitting right in front of you and you have to be able to look around it, watch it working, poke at it, and take it apart -- long after you've been exposed to enough information about it to build a model of it. Why? One receives several answers to this sort of question, but my favorite (and also the most common) is what I call the *gratuitous deficit* response. For example, Maybe you build a model but it decays. Maybe there isn't enough capacity. Look what we have here. We have a theory that makes everything easy by postulating explicit encodings of every salient fact in the world. And then we have this sort of lame excuse that pops up to disable the theory when anyone notices its blatant empirical falsehood, without at the same time admitting that anything is fundamentally wrong with it. Maybe the computational complexity of reasoning with realistic world models is trying to tell us something about why it's harder to think about a derailleur in the next room. What exactly? It's hard to say. But that's to be expected. Expecting it to be easy is a sign of addiction to the easy answers of the orbiculi.