2. Fundamentos análisis multivariante
2.4. Anosim
The stereotypical movie robot is an android with a voice-recognition in- terface and a sardonic sense of humor. In fact the most common type of robot is quite a bit less exotic than that, often built out of Legos, with a couple of motors and some big cables that need to be held up by a gradu- ate student who follows the robot around the room. It will be worth- while for us to sidetrack for a moment and talk about robot intelligence, since some of the most important work in swarms and swarm intelli- gence is being conducted in that field.
In the early days of electronic computing, robots were a perfect labo- ratory for the implementation of artificial intelligence, and the classic ro- bots were based on the paradigm that has come to be known as GOFAI (Good Old-Fashioned Artificial Intelligence), which is practically synon- ymous with the symbol-processing paradigm. Robot intelligence was presumed to operate—in fact, human intelligence was presumed to oper- ate—through the manipulation of symbols. Presumably, the same rules of logic should operate on any set of symbols, no matter what the subject domain. Generally, a robot or human mind was presumed to contain a
central executiveprocessor; this was presumed to include a kind ofinfer-
ence enginethat operates on data that have been prepared by a perceptual
system, and the executive’s output was sent to effectors, or motor de- vices, which were able to act upon the world.
Given the assumptions of cognitive science and the profound vanity of humans generally, it is not surprising that scientists would assign the will (central executive) to a central role in an artificial mind and imple- ment a system of information processing that followed directly from the classical academic philosophies of the universities, and before that, from the pronouncements of the Church. The early researchers had no way of
knowing that scientific psychology would fail to support their intuitions about human nature and mind—remember that, before the “cognitive revolution” of the 1960s and 1970s, almost all scientific psychology was behavioristic, dismissing everything mental as unworthy of scientific study. Thus there was not a well-developed understanding of things cognitive.
In robotics, it turned out that the symbol-processing approach just didn’t work out very well. As the robot’s world became more compli- cated, the robot’s mind became more complicated; it had more symbols to retain. Worse, maintenance of complicated chains of logic meant that it had to figure out that when one fact changes, other facts have to change, too. It is not trivial to deduce which ones. For instance, if you are told that someone has just turned out the lights,youknow immediately that they can’t see where they’re going. A symbol processor has to figure it out. GOFAI robots might stand in one spot for 10 minutes contemplat- ing their next step, trying to figure out whether they would bump into something.
In a highly readable paper called “Elephants don’t play chess,” MIT researcher Rodney Brooks (1991) proposed an alternative to GOFAI (a tradition with deep roots at MIT, by the way). Brooks argued that sym- bols must begroundedin a physical reality if they are to have any mean- ing. Brooks’ view solved some computational problems that had made real-world robotics intractable, and the successful results seen in Brooks’ laboratory lent support to his conjecture.
Brooks’ robots’ minds are organized according to a principle called
thesubsumption architecture.A subsumption architecture is built from the
bottom up; simple robot behaviors are developed, and then these are fitted loosely together. They remain independent of one another, each module doing its specific part without consulting the other. For instance, Brooks’ soda-can-finding robot, Herbert, has a collision avoidance mod- ule that sends a signal to turn it when it comes too close to something it might bump into, and a can-pick-up module to pick up the cans when it finds them. Herbert can pick up cans whether it is or is not avoiding anything, and can avoid things without picking up cans, or do neither, or do both. The modules are independent. Further, both modules are grounded in the world; that is, avoidance happens when there is some- thing in the world to avoid, and picking up happens when a can is de- tected in the world. There is nothing like a central executive control.
The quotable Brooks has said, “The world is its own best model. It is always exactly up to date. It always contains every detail there is to be known.” Rather than maintaining a symbolic representation of the world, Brooks’ robots simply respond to the world as it presents itself to
their sensors. The effect of this is that computation can be reduced to a tiny fraction, perhaps 1 or 2 percent, of what was required by previous robots, programs can be smaller, storage and memory requirements are reduced, and as a result smaller, smarter robots can be built more cheaply.
Brooks’ robots are able to navigate messy scenes. For instance, Her- bert roams around cluttered laboratory offices looking for empty cans, which it picks up and throws away. No set of rules could prepare a sym- bol processor for all the kinds of complexities that might exist there, for instance, tipped-over chairs, wadded-up papers—tipped-over gradu- ate students, for that matter. The bottom-up subsumption architecture, though, is able to adapt to whatever conditions it finds.
Tom and Jerry were two identical robots that were programmed to in- teract with one another. They were given an urge to wander about, that is, to perform random actions, a repulsive force that kept them from bumping into things, and were programmed as well with an attractive force, to make them move toward distant things. They also had a motion detector sense that allowed them to detect moving objects—most likely one another—and follow them: a possible genesis of robot sociality.
Construction of subsumption architectures is bottom up: it is incre- mental; small parts are constructed and then fitted together in a way that Brooks likens to evolutionary change. Communication among various modules is minimal, with tight coupling between modules and the world. The hallmark of such an intelligence is that it is decentralized, a theme that should be getting familiar by now. In the subsumption archi- tecture, behaviors are separate, each one defined to accomplish one small, distinct goal, yet someone watching the robot feels that there is an intentional unity about the way it moves. Wheels are programmed to stop when the robot approaches a possible soda can; arms are pro- grammed to reach up when the wheels stop turning; a hand is pro- grammed to grasp when the light beam between fingers and thumb is broken—an observer sees the robot go to the table, reach out, and pick up a can. Oddly, goal-directed behavior emerges, or so it seems, from a se- quence of fixed action patterns linking perceptions to actions. Pushes look like pulls through the prism of our biological brains.
Brooks’ model brings up a new facet of this swarm business. We note that the MIT robots areautonomous:they are self-contained and self-suf- ficient—with an occasional human helping hand. But inside the robot’s mind there is something of a swarm of modules, you might say, a multi- tude of subroutines that perform particular functions.
makes it clear that he is talking about robots, man-made minds, and sim- ple ones at that, Minsky takes the model of AI modularity and tries to use it to explain the workings of the human mind. Minsky’s coffee-table book,Society of Mind,contains 270 essays, each one packaged as a pearl of wisdom or aphoristic reading on a psychological or cognitive topic. The theme of the book is that mind is composed of a “society” of specialized functions or “agents” that operate more or less independently of one an- other. Unfortunately, the subsumption approach that is so successful with robots has no basis in psychology—though there is evidence of modularity in the organization of the brain, the metaphor of a society of cognitive agents is exactly the kind of reified fiction that scientific psy- chology rejects. There is no homunculus inside the head, never mind teams of homunculi.
Near the end of his book, Minsky summarizes with the statement: “Minds are simply what brains do.” The statement is concise and super- ficially plausible. It seems to do away with the mind-body problem in one slicing phrase and reduces much of Western philosophy to dry chat- tering. But the statement is obviously incorrect. There is doubtless a relation between brains and minds, but there are very many brains that don’t do minds at all, for instance, bird brains, and bug brains, and sleeping brains, and fetus brains, and damaged brains—dead brains definitely don’t do minds. There are also lots of things that brains do that are not mind: having seizures, controlling digestion, growing, me- tabolizing oxygen, hemorrhaging. Further, there doesn’t seem to be any reason that minds require brains—we can imagine aliens having minds, and there is a strong enough case for computers having minds, at least potentially, that Minsky’s statement just can’t be true in an absolute sense.
Even worse, the statement—and the cognitivistic philosophy that supports it—assumes that a brain by itself, we’ll even grant that it is a healthy human brain, would develop into a mind, that it would become conscious and learn to think (which we take to be minimum require- ments for something to be called a “mind.”). This ignores all social sci- ence research for the past hundred years. In order for a brain to become mental, it needs to have interaction with other minds. A fetal brain grown to adulthood in a vat containing nutritious fluids would never ac- quire the qualities of mind. If it were transplanted into a Frankensteinian body, it would be unable to reason, remember, categorize, communicate, or do any of the other things that a minimal mind should be able to do. Feral humans, that is, humans who have grown up in the wild, have never been known to show signs of having anything like a mind—it is not sufficient to have perceptual stimuli; social ones are necessary. The
relation between mind and society is unquestionable, but is ignored by Minsky’s oversimplifying, where the “society” is moved into the cavern of the skull.
We would have thought the issue was more or less settled in the research debate between Ernest Hilgard, a cognitivist, and Nicholas Spanos, a social psychologist, in a fascinating series of investigations into the nature of hypnotic responding, multiple personalities, and other forms of psychopathology. Hilgard’s neodissociation theory (e.g., 1979, 1977) postulated autonomous cognitive processes inside the mind, which under certain circumstances separated and acted independently from the self. For instance, in hypnosis, cognitive subsystems could come under the control of a hypnotist. Hilgard’s mechanistic model re- sembled Minsky’s society of mind, with agentic substructures expressing autonomous volition beyond the reach of the individual’s self. Spanos (e.g., 1982, 1986), on the other hand, argued that individuals were sim- ply acting in ways they believed were appropriate for situations labeled “hypnosis,” “multiple personality,” and so on. Anyone can act like they have multiple personalities, but only a few people convince themselves. According to Spanos’ sociocognitive theory, hypnosis and certain forms of pathological responding were manifestations of normal behavior; no internal systems or “agents” were invoked to explain behaviors. Hypno- sis, in this view, is something the person does—not something that hap- pens to him. Where Hilgard’s experiments would seem to show that a hypnotist could communicate with different autonomous subsystems of the hypnotized person, for instance, a “hidden observer” within the sub- ject, sealed off from the rest of the cognitive system, Spanos would repli- cate the experiments under slightly different conditions to show that the hypnotized individuals were just doing what they thought they should be doing. The neodissociationists’ internal special processes were shown to be fictitious, theoretical constructions with no basis in reality. Again and again Hilgard and his colleagues attempted to prove the existence of cognitive substructures, and repeatedly Spanos’ case for strategic role enactment prevailed. Nothing is gained in thinking of the mind as a “society” of interacting agents.
There are some differences between multiagent systems such as the subsumption architecture and swarms such as are being described here. In swarm intelligence the global behavior of the swarm is an emergent ef- fect of the local interactions of swarm members; in a system of multiple autonomous agents, however, the system’s behavior is more nearly asum of the agent’s contributions. The effect may strike a human observer as being surprisingly well choreographed, as if the robot were pursuing a
behaviors are exactly what they have been programmed to be, and the cause of the surprise is anthrocentrism, not emergence. Further, individ- uals in a swarm are usually relatively homogeneous, while “society of mind” or subsumption modules are assigned to specialized tasks. One ant is the same as the next; one bee does pretty much the same thing as the next bee. Grasping is very different from looking, which is different from rolling.
Rodney Brooks has expressed the opinion that robots will continue to get smaller and smaller; indeed the simplicity of the subsumption archi- tecture enables such a miniaturization. For instance, he says, imagine a colony of “gnat robots,” living in your television, that come out when the set is off and dust the screen. Indeed, the stage is set for such develop- ments; it appears likely, though, that the miniature robots of the future will probably not be seen as autonomous agents, each one fully empow- ered to accomplish whole tasks. The microrobots of the future will be swarms.
On an even smaller scale are the nanotechnology robot swarms envi- sioned and described by Kurzweil (1999). Large assemblies of the constit- uent nanomachines will be able to create any desired environment. One example of nanorobot swarms is described by Hall (1994, 1995), in which each nanorobot, called a Foglet, is a cell-sized device with 12 arms pointing every which way, each with a gripper. A large assemblage of these devices can not only cooperate to form large structures, but also form distributed intelligence. But let’s come back to the present for the moment.
Toshio Fukuda’s laboratory in Nagoya, Japan, is a day-and-night bus- tle of graduate students, postdocs, assistant professors, and engineers, planning, designing, and building robots of many types. For Honda Cor- poration, Fukuda has built a life-size android robot with two legs, two arms, a trunk, and a head, balancing and walking with a humanlike gait. If you shove this robot backwards, it will flail its arms and catch its bal- ance just like a person would. It’s a lucky thing, too, since it weighs nearly 600 pounds. Fukuda’s lab also produces robotic gibbons, two me- chanical arms and the trunk of a body; a gibbon is able to train itself to swing from limb to limb in an artificial jungle with both overhanded and underhanded brachiation movements. When a gibbon’s hand is knocked from its branch, the robot swings incrementally higher and higher until it can reach the branch again—just as a real gibbon or mon- key would do. In lectures, Fukuda likes to show a video of a gibbon in a jungle setting and say, “That’s not a robot, that’s a real monkey.” In the next scene a robotic gibbon is depicted, with uncanny resemblance to the live one. Fukuda says, “That’s a robot” (Fukuda, 1998).
Fukuda’s lab is developing robot swarms. Where Brooks’ subsump- tion methodology took the cognitive executive control out of the pic- ture, decentralizing the individual robot’s behavior, Fukuda’s robots re- place the individual’s control with reflex and reactive influence of one robot by another. These decentralized systems have the advantages that the task load can be distributed among a number of workers, the design of each individual can be much simpler than the design of a fully auton- omous robot, and the processor required—and the code that runs on it— can be small and inexpensive. Further, individuals are exchangeable. We think back to 1994 when the NASA robot named Dante II ambled several hundred feet down into the inferno of Mt. Spurr, a violent Alaskan vol- cano, to measure gases that were being released from the bowels of the earth. The 1,700-pound Dante II toppled into those self-same bowels and would now be just an expensive piece of molten litter if helicopters bear- ing earthlings had not saved it—for instance, if the accident had hap- pened on another planet, which was what Dante was being developed for. If, instead, a swarm of cheap robots had been sent into the crater, per- haps a few would have been lost, but the mission could have continued and succeeded. (Dante II was named after Dante I, whose tether snapped only 21 feet down into Mount Erebus in 1993—another potentially ex- pensive loss.)
Just as knowledge is distributed through the connections of a neural network, so cellular robots, as they are called, might be able to encode representations through their population, communicating as they in- vestigate diverse regions of the landscape—a real, three-dimensional landscape this time. The concept of cellular robotics, as Fukuda tells it, is based on the concepts found in E. O. Wilson’s research on insect socie- ties, where global intelligence emerges from the local interactions of