I promised earlier to take up the Humpty-Dumpty question: how do you put the pieces together? From a lump of problem-solver, a bag of mixed pattern recognizers, a language translator, an associative memory, etc., how do you create a coherent intelligence? And a little more: how do you give it a sense of purpose? How do you build in emotions, awareness, etc.?
As already indicated, there has been more work than progress on these matters. Some of the work is worth mentioning, however, for it involves actual robot-building projects at Stanford and M.I.T. If the first generation of robots is not already with us, we are at least seeing the first generation of robot embryos.
Before getting into robot projects, it deserves mention that GPS (the "General Problem Solver" of Newell, Shaw, and Simon) had one answer to the question of purposeful behavior. Essentially, it identified purposefulness with a pushdown list. (That's one of those growing lists of things that have been put aside while tending to other things -- some of which have wound up on the list themselves.) In principle, GPS was versatile but single-minded: at any given moment, it had only the purpose dictated to it by whoever was feeding it a problem. It could digress, however, to take up other problems whose solution looked useful towards the solution of the main one. It could digress from the subgoal in turn, for the same sort of reason. The list of deferred business could get as long as storage would allow, but eventually the system would work its way back from each sub-problem to the one that suggested it, and finally to the main problem -- if only to report failure. In this sense it was (grimly) purposeful.
A bit of doggedness is necessary in a system which is to be more self-willed than, say, a falling leaf. The systemic thoroughness of GPS, however, seems somehow to overdo it. (Doesn't a purposeful creature ever forget what it was up to?) In any case, perseverance through a series of deliberate digressions is only one aspect of goal-directed behavior. What about consciousness of purpose? What about the invention of a purpose to start with? What about purpose-seeking behavior?
Ironically, the slavishness of GPS may result not from the fact that it is given its purpose by a programmer, but that it is given only one purpose at a time. A moment's reflection should confirm that human beings, too, are handed their more primitive purposes from on high, in the form of built-in drives. They are handed, however, purposes -- not "a purpose." (Were it otherwise, how could we hear so much talk of "finding a purpose in life"?) And that is just where the fun begins.
The all-important point is that a man's purposes can conflict: there are trade-offs and fringe benefits in all directions. Those drives which have been most thwarted at a given moment speak loudest. Any one drive can be overridden on behalf of any other. The sense of being "self-willed" results not from an absence of built-in directives, but from the plurality of such, and the constant necessity of arbitrating their demands. (It is doubtful that we could rise above blobhood if we had to invent for ourselves the will to survive, the impulse to eat, the urge to merge, and all that sort of thing.)
A recent model which is more satisfying in some of these respects is a program called "ADROIT", by Leonard Friedman.(6) This program starts squarely from the premises just mentioned: that an organism has a multiplicity of built-in drives, that these compete for servicing, and that the squeakiest wheel gets the grease. More precisely, drives (in the model) are implemented by "releasing mechanisms" which are responsive to certain signals in the environment -- changes in the body which are the natural consequences of neglecting the drive, perception of something nearby that would satisfy the want, etc. The most stimulated "releasing mechanism" emits the strongest signals and gets the go-ahead to release an appropriate response in the organism's repertoire.(7)
There are multiple layers of "releasing mechanisms" in ADROIT, and the signals which they emit (when they are given the go-ahead) can be fed to other releasing mechanisms as input, thus offering unlimited opportunities for hierarchical organization and feedback.
Bear in mind that ADROIT is only a simulation. It dodges all the hard technical questions, focusing exclusively on the logical question of control. Both perceptions and appropriate "effector" routines are taken for granted. (For example, if "nest building" is the drive, then "recognition of a suitable site" is assumed as a predigested input to the releasing mechanism, and "hunting for grass and twigs" is assumed as a ready-made response pattern that can be activated.
For this reason, it is only as meaningful as the choice of "drives," "signals from the environment," and "responses" to be represented, and the way they are interconnected. A naive choice might produce a model that resembles a city council, its actions governed by the last group which has been in to protest. (You might call it "demonstration-driven.")
No doubt there is a streak of this in actual behavior: the most resolute toiler must undergo a schizophrenic change of mood eventually, as his stomach begins to growl or his body gets ready for sleep. One would like, however, to couple this lifelike subjection to more than one demanding master with an equally lifelike ability to start a project, set it aside, and come back to it as often as need be to finish it.
There is nothing in ADROIT either to hinder or facilitate this -- it's a matter of how you use the logical entities provided by the simulator. However, ADROIT is predisposed to wired-in plans that rely on the environment to provide the next cue. To simulate more cerebral activity, it would have to take on the dodged details of how "responses" (other than simple reflexes) actually work.
Now, as to robot projects. There are two of these, and you might count a couple of hand-eye combinations as some fraction of another. There is a "hand-eye" project at Stanford University, for example; the "hand" stacks children's blocks with the help of the "eye." The Japanese firm of Hitachi advertises a similar rig, and envisions it working some day on an assembly line. The full-scale (but primitive) robots are at Stanford Research Institute and the M.I.T. Instrumentation Lab. The one at S.R.I. does things like pushing big blocks into a heap at the center of the room; the one at M.I.T. is supposed to make like a tourist some day on the surface of Mars.
The Stanford robot contrasts sharply with the ADROIT model (which was one reason for mentioning the latter). Like GPS, the robot is slavish: Master barks an order, and until Master barks another, that one order is Robot's whole reason for being. On the other hand, the robot is made to function in a real world. Make-believe capacities are not allowed, that's a ground rule. And the nature of the problems which are set before the robot require that a variety of traditional "artificial intelligence" capacities be brought together into a working whole.
The publicity blurb for the S.R.I. robot is quite explicit about one thing: pulling artificial intelligence efforts together is a major aim of the project, not just a by-product. In practice, all they've pulled together so far are a line-recognizing program, a scene analyzer, and a theorem prover used as a planning program.
Much of the preliminary effort has gone into necessary but undramatic underpinnings -- the robot hardware, communications between it and the computer, system programming, and other mundane matters.
Here's how it works.(8) The robot is a little motorized cart. For a head it has a television camera, which it can nod up and down. It also has feelers which tell it when it is bumping into something. It can turn its drive wheels independently, allowing it to go forward, go backward, or pivot. The "brain" is lodged separately in a computer; you might say that its spinal cord is an airwave, for information passes between "body" and "brain" by wireless. (Sorry about that; current-generation computers are too big to trundle around on a cart.)
Commands are entered into the computer via a terminal. They have to be carefully phrased, and must fall within the robot's bag of tricks. Say the instructions are to collect a bunch of blocks at the center of the room. (You'd better say that, because the robot doesn't really have any other tricks to speak of.) The robot gets its bearings: it takes pictures and sends them back to the computer. (Even this much is a major undertaking. We do things like turning our heads to get a look so effortlessly that it takes a robot project to make us appreciate our gifts.) The "scene analyzer" program looks at lines and corners (after a lower-level program has detected the lines), then elaborates them into objects. It helps that the objects are known to be cubes.
The theorem prover then attempts to prove that, starting from where they are, the blocks can be moved to the spot requested. The proof is in the form of a series of moves the robot could make to get them there. If a proof is found, it then serves as a plan of action, and is carried out. All of which sounds pretty simple, but it isn't. There are endless little mechanical details to tend to, and the "moves" with which the theorem prover must work might best be described as fits and starts. (They have to be chopped pretty fine, since maneuvers such as pivoting are carried out in increments: "Turn your right wheel forward 5 notches while the left goes backward 5." If the notches were big, it would be difficult to execute delicate movements.)
With rare flare, S.R.I. has produced a promotional movie on "Shaky, the Robot." The nickname is well chosen: Shaky is the very epitome of a jerking hulk. When he approaches a block, for example, with intent to push it, he jerks to a halt, edges forward gingerly, contacts it with his feelers, stops again, backs off, and then abruptly heaves to. (At each step, the computer is processing interrupts, reviewing progress, etc.)
Both Shaky and a hand-eye combination developed earlier at M.I.T. remind one of Stephen Potter's formula for winning at ping-pong. You call the opponent's attention to his elbow: in the friendliest possible manner, you advise him that it should follow through like so (you guide it through a motion). Once he's thinking about his motions, they fall to pieces. Shaky thinks about his motions -- detail by detail.
So far, just a clever combination of line detector, scene analyzer, and theorem prover -- plus a whole lot of mechanical engineering and system programming. Work is in progress, however, to integrate these with a natural language processor, a voice recognition unit, and a whole lot more -- not to mention upgrading of the components already in place.
Yet another manner of beast is the M.I.T. robot.(9) This one, still on the drawing boards, takes its inspiration from physiological theories, á la Warren McCulloch. Like the Stanford robot, it is to be a wheeled vehicle, with television "eyes"; its logical make-up, though, veers strongly in the direction of ADROIT.
MR (for "Mars Rover") has a mission: get out and explore Mars.(10) Send pictures to the folks back home. ("Having a wonderful time, wish you were here.") Transmission across 35 million miles or more is difficult and expensive, so the robot should summarize the scene a little before transmitting it. It should use its head a little as to what would make an interesting shot. And it should avoid toppling into the nearest ravine.
MR is designed to get by on a minimum of overt commands; a pioneer must be self-reliant. It isn't told about its mission, things just work out that way if the robot responds as expected to certain built-in itches. The itches have been contrived so that the environment should trip off useful responses.
These modes of behavior have been suggested for the robot:
Each mode of behavior is associated with various cues from the environment (not excluding direct commands from Earth). A "command computer" selects the mode of behavior at any moment on the basis of all the cues available to it just then. The decision mechanism is exactly like that in ADROIT: squeaky wheel gets the grease.
Whereas Shaky is run by a single, general-purpose computer, MR will carry around an assembly of special-purpose computers: one for preprocessing visual signals, one for calculating range and direction, one for controlling motion, one for timing, one for choosing what to do next, one for thinking about it all, and so forth. (The various computers correspond, supposedly, to various parts of the brain.)
Early models of MR will not attempt to plan, learn, or remember anything. Like ADROIT, MR will simply have wired-in responses which are hopefully adequate to their purposes; the environment does the robot's thinking for it.
The grand design for MR is well in advance of the implementation. There will be successive models for successive Mars missions. The various boxes are in various states of development; some are already wrought in special-purpose hardware, some are simulated in a general-purpose computer, some are rather loosely specified.
Because of the robot's mission, the visual equipment has been brought forward first. (The first robot on Mars will probably have little more than this.) Where simulation has been employed, it is what might be called "hard" simulation: that is, there are definite physical realizations in mind for the things simulated. (There are degrees and degrees of make-believe in computer simulation.)
MR is no match for Shaky in such matters as planning or comprehending human orders. Even its visual apparatus, a specialty of the house, seems a little primitive in comparison with the scene-analyzing program in Shaky. (MR is one-up, though, in having binocular vision.) On the other hand, MR seems more natural and lifelike. Also, it is quite modular; as additional pieces of the grand design are brought to life, it should become more and more impressive. The use of special-purpose computers should save a lot of attention to system programming -- provided their respective functions turn out to be well-chosen.
A point worth noting is that MR, if it realizes its authors' hopes, will be the first artificial intelligence with something like a sense of curiosity. Its "chart the environment" urge resembles the instructions built into a nine-month old baby: "When you've got nothing else to do, crawl around and look for an unfamiliar object. If you find one, pop it in your mouth."
Where are we, then? In one shop, theorem-proving has been integrated with scene analysis; language processing and speech recognition are on their way into the mix. In another shop, the apparent organization of the vertebrate nervous system has been imitated; conflicting drives, responding to cues in the environment, have been used to provide direction and control.
These statements have to be qualified with the observation that the pattern recognizers, theorem provers, built-in-drives, and everything else in question are primitive and limited. The capacities actually realized by MR, for instance, are limited to the preparation of line drawings with shading.
The Stanford robot, as one critic has jibed, lives in an environment that would drive a human being insane. (It's a single room with bright walls, a bright floor, and a black stripe where wall meets floor. Sometimes there are big blocks to play with.) The MR robot will face a real enough environment, but is expected to get by in it only because it has such simple chores to do, and because nothing will ever force it to flee for its life.
One is left with the sense that while there are glimmerings of self-will in the design of MR, and glimmerings of connected thought in Shaky, neither is at all satisfying as to awareness. Neither is very sophisticated about learning anything. The signals received by the "command computer" in MR, which it weighs in deciding what to do next, may be compared with emotions -- but intuitively, one can't imagine that they feel like emotions.
An early enthusiast of robots, now turned critical, points out that none of the robots proposed so far has had any knowledge about its knowledge.(11) There is no self-image, no concept of a concept. It is easy enough to say that this can be provided by adding a model of the robot itself to the other models stored in its memory -- it's been said lots of times --; but nobody has ever detailed such a scheme. Ernst is probably right that nobody knows how.
Ending on such a note, it would be easy to conclude that my argument has more or less disposed of itself; robots aren't much to howl about. But remember, I undertook to cite work in progress, not miracles perfected. All things must begin with beginnings; robots are having theirs.; Robotry is in at least as good shape today as was rocketry in the thirties. What is important is not the shortcomings and vexations of a moment in time, but the obvious world-event that is shaping up.
|
|
BACKTALK FROM READERS |
| How Shall I Compare Thee? Biology Gets a Metaphor. |
 |
 |