Life Insurance: Algorithms and Heuristics
Dateline: August 30, 1997
This article (and next week's, and the one after) is based on my reading of Daniel Dennett's 1995 book, Darwin's Dangerous Idea: Evolution and the Meanings of Life, with a few of my own thoughts thrown in.
Prejudices, even subconscious ones, tend to make us draw conclusions that may not be justified by the evidence concerning their object. Says Dennett, "We are all apt to do this, especially when we try to construe in simple terms the take-home message of work outside our own field. We tend to read, with bold brushstrokes, what we want to find." Caveat lector, therefore, for you are in double jeopardy: you may read into what I write more than I intended, and what I write is likely to reflect my own misreading of what Dennett and others have written.
Does it matter? In the final analysis, probably not. It seems to me our algorithm for reading/listening/paying attention to what others say, write, and do is a heuristic one. To read Dennett in exactly the way he intended, you would have to be Dennett and in his exact frame of mind at the time he wrote (and you read) every word. If such were possible, the human population would consist entirely of Dennett clones and we could all just as readily write as read his works! Jorge Luis Borges' short story, Pierre Menard: Author of the Quixote, captures this notion.
AN algorithm can be described as something which:
1. Can be recorded in any medium or set of symbols,
2. Requires no intelligence to operate, and
3. Is guaranteed to produce a result.
A key part of Dennett's central thesis is that the process of evolution (whose "Modern Synthesis," the version of evolutionary theory accepted by most biologists today, combines Darwinian natural selection with Mendelian genetics) is nothing more than the mechanical, automatic operation of one huge (and growing) set of algorithms.
If that is so, then from the above features we can say that evolution itself can be expressed (can happen) in any medium—biological, chemical, physical, or symbolic; it is self-actualizing—an intelligent controller/operator/button-pusher is not necessary (but, in my view, is not precluded, either—see footnote); and finally the outcome of evolution—whatever that outcome may be—is certain.
The algorithms of natural selection, says Dennett, do not predict any given species, and the reason for that is that chance (in the form, for example, of natural disaster such as the asteroid that put paid to the algorithm governing dinosaur evolution) is present at each step of the algorithm.
The "heart of the power" of Darwinism, for Dennett, is this: "An impersonal, unreflecting, robotic, mindless little scrap of molecular machinery is the ultimate basis of all the agency [chemical agents in cells], and hence meaning, and hence consciousness, in the universe."
Plato had a sense of this, and expressed it in his analogy between a republic and a person—both being Organizations composed of smaller, interacting, more or less mindless agencies. (Mindless, that is, compared to the knowledge and awareness of the Organization.) Dennett comes to the same conclusion through his study of Darwinian processes in biology: "That is what a mind is—not a miracle machine, but a huge, semi-designed, self-redesigning amalgam of smaller machines, each with its own design history, each playing its own role in the `economy of the soul.'" This view is essentially that of Marvin Minsky's Society of Mind.
Why do these agents/machines/algorithms cooperate, as they must if they are to combine to form a viable organism? Game Theory, which we came across in a previous article and which was John von Neumann's focus before his death, suggests an answer, and it may be no coincidence that Game Theory is widely applied in both Darwinian (biological) evolutionary research and in Alife (subject of yet another previous article!)
The theory figures in John Maynard Smith's Evolutionary Stable Strategy (ESS) theory, which says in essence that the most successful evolutionary strategy or trait (or algorithm) will become dominant in any given population. Having eliminated weaker competing strategies, the ESS will then tend to compete against the many copies of itself which must, by definition, abound in the population. This seems to me to be demonstrably true of Homo sapiens which, having hit upon the winning strategy in the game against gorillas, lions, tigers, wolves, and bears turned in upon itself in a competition that continues, albeit in a less physical and less overtly vicious form, today.
Evolutionary subroutines (yet another name for machines, algorithms, agents) are bound by what Dennett calls "the notorious need-to-know principle of espionage: give each agent as little information as will suffice for it to accomplish its share of the mission." It seems to me that as "agents" of the Noosphere (you guessed it—see previous articles), and/or of Machina sapiens, then perhaps we too can never be privy to the Big Picture of Life, the Universe, and Everything. (Dennett, who does not like Teilhard de Chardin or his Noosphere, would disagree.)
It is certain that our individual brains have limited storage and processing capacity (yep, another previous article), and our sense organs are limited in what they can sense. But even though we will be weaker than Machina sapiens in these regards, will our ability to communicate with it make up for our deficiencies, just as today we can communicate with an alpha-proton spectrometer on Mars and make sense of what it sees, even though we have no way of seeing it directly?
Or will such vicarious experience leave something out?
In the changeover from analog electromechanics to digital computers that has been taking place in airliner cockpits during the 1990s, designers of the European Airbus replaced the pilot’s yoke with a small joystick very much like those used for playing computer games. Through the old mechanical yoke, said Lufthansa's chief pilot, a pilot could "feel" every part of the airplane and quite literally sense (through tactile feedback) when something was wrong. With the new electronic joystick, the pilot lost that close coupling with the aircraft and was forced to rely on the computer's opinion of the state of the plane.
The Airbus designers seemed to throw out a set of meticulously engineered algorithms, honed to near-perfection through millions of years of evolution, in favor of some pasty-faced, sleep-deprived, 20-year-old programmer’s notion of what a flight control algorithm ought to be! (Just kidding, programmer friends!)
Whether one system (joystick or yoke) turns out to make for safer airplanes than the other is not the point here. The point is that if the pilot were to need, or would benefit from, tactile feedback, he or she would be out of luck with a joystick. While we can easily bestow tactile feedback (by re-installing the traditional yoke, as U.S. aircraft makers did), we cannot bestow true, original feedback for a sensation we are not biologically equipped to handle.
As Dennett says, "when the complexity of encountered environments arises,... and unpredictability becomes a more severe problem, a different design principle kicks in: the commando-team principle, illustrated by such films as The Guns of Navarone: give each agent as much knowledge about the total project as possible, so that the team has a chance of ad-libbing appropriately when unanticipated obstacles arise."
I propose that this principle kicks in when, for example, a pilot faces a sudden crisis and there is literally not enough time to think. There may only be time to react, and the reaction must be close to instantaneous. The hands on the yoke, through continual but slightly time-delayed feedback from the brain about the state of the aircraft and its environment in general, are almost—but not quite—ready for any emergency. In an abrupt emergency, the hands must be ready to ad lib in an instant to avert disaster. For the duration of that instant, the hands perhaps literally have a mind of their own.
"Once we get to Popperian creatures," says Dennett, "creatures whose brains have the potential to be shaped into inner environments with pre-selective prowess, what happens next? How does new information about the outer environment get incorporated into these brains? This is where earlier design decisions come back to haunt—to constrain—the designer. In particular, choices that evolution has already made between need-to-know and commando-team now put major constraints on the options for design improvement."
"We engage in our share of rather mindless routine behavior, but our important acts are often directed on the world with incredible cunning, composing projects exquisitely designed under the influence of vast libraries of information about the world."
Heuristics
"One might even define the task of the field of AI as the creation and investigation of heuristic algorithms," says Dennett. The difference between an ordinary and a heuristic algorithm is that the latter has an inherent ability to produce a much broader array of possible outcomes, and is therefore riskier.
He adds: "Mother Nature has never aspired to absolute certainty; a good risk is enough for her." Furthermore, "there are risky, heuristic algorithms for human intelligence in general.... Here is where Penrose [mathematician Sir Roger Penrose, author of The Emperor's New Mind, an AI-doubting book we discussed in a previous article] made his big mistake: he ignored this set of possible algorithms—the only set of algorithms that AI has ever concerned itself with—and concentrated on the set of algorithms that Gödel's Theorem actually tells us something about." (I, too, noted in my earlier article the absence of any discussion of heuristics in The Emperor's New Mind. Dennett informs us that "In the wake of the commentary his book provoked, Penrose now grants that heuristic programs are algorithms as well....")
Mathematician Kurt Gödel's famous Incompleteness Theorem seems to militate against algorithms, even heuristic ones, as the underlying explanation for everything. Says Dennett, "... Gödel, anchored by Turing to the world of computers, tells us: every computer that is a consistent truth-of-arithmetic-prover has an Achilles' heel, a truth it can never prove, even if it runs till Doomsday. But so what? ... Gödel himself thought that the implication of his theorem was that human beings—at least the mathematicians among us—cannot, then, be just machines, because they can do things no machine could do. More pointedly, at least some part of such a human being cannot be a mere machine, or even a huge collection of gadgets.... Mathematicians' minds cannot be their brains, Gödel thought, since mathematicians' minds can do something that no mere computing machine can do.... they can just see that certain propositions of arithmetic are true. The idea would be that they don't need to rely on grubby algorithms to generate their mathematical knowledge, since they have a talent for grasping mathematical truth that transcends algorithmic processes altogether."
But, says Dennett, some algorithms are "pretty darn good at playing the Turing Test or imitation game. In fact," he continues, "there is one actual one on my Toshiba, a stripped-down version of Joseph Weizenbaum's famous Eliza program, and I have seen it fool uninitiated people into concluding, like Edgar Allan Poe [who famously unmasked a bogus 19th century chess-playing automaton], that there must be a human being issuing the answers."
Even attributes as slippery as Truth and Beauty have an algorithmic base: "... underlying our general capacity to deal with such `ingredients' [of truth and meaning] is a heuristic program of mind-boggling complexity. Such a complicated algorithm would approximate the competence of the perfect understander, and be `invisible' to its beneficiary. Whenever we say we solved some problem `by intuition,' all that really means is we don't know how we solve it. The simplest way of modeling `intuition' in a computer is simply denying the computer program any access to its own inner workings. Whenever it solves a problem, and you ask it how it solved the problem, it should respond: `I don't know; it just came to me by intuition.'"
The algorithmic evolutionary perspective is gaining ground to the extent that some evolutionists are "intent... on replacing the `Standard Social Science Model' with a properly Darwinian model of the mind." To me, this is another sign of a scientific revolution in the making. Dennett is more cautious, and advances his own, only slight, revision of the standard model:
"Whereas animals are rigidly controlled by their biology, human behavior is largely determined by culture, a largely autonomous system of symbols and values, growing from a biological base, but growing indefinitely away from it. Able to overpower or escape biological constraints in most regards, cultures can vary from one another enough so that important portions of the variance are thereby explained.... Learning is not a general-purpose process, but human beings have so many special-purpose gadgets, and learn to harness them with such versatility, that learning often can be treated as if it were an entirely medium-neutral and content-neutral gift of non-stupidity."
We’ll deal with cultural evolution and "memes" next week. The week after, all being well, we'll look at perhaps the slipperiest algorithm of them all: Morality.
Until
next week,
FOOTNOTE: Dennett takes several of the world's great thinkers to task for failing to see or to accept the mindlessness of creation. He rejects philosopher David Hume's eminently reasonable conclusion that "A total suspension of judgment is here our only reasonable recourse," and himself sits in heavy judgment on all who would so much as hint at the possibility of an Intelligent Designer. I think this is unfortunate and smacks of the fanaticism he condemns in others. Unless and until science can tell us with assurance the full story of the origin of the universe (and one modern mathematical physicist, Frank Tipler, has tried—and arrived at an Intelligent Designer!) then I think Dennett should follow Hume's example.
NEXT WEEK: Life of Riley: Memetics and Cultural Evolution.
READER COMMENTS/REQUESTS:
First, a request:
Hi, i'm currently trying to implement a java chess program which includes a computer opponent. Unfortunately, I'm having a difficult time finding any good books that explain, for example, the minmax and alpha-beta pruning at an undergraduate level. Would you be able to help me out?
Uriel Avalos (ua697f94@panther.adelphi.edu)
I'm not a programmer, but hopefully someone out there is and can help Uriel out. Please respond to him direct, or write to me and I'll pass it on. Thanks.
… and a comment:
Edward Greaves (egreaves@cybernex.net) wrote: Not sure WHO gets this info, but I'll write it in two parts. General: So far this page has been well maintained, and enjoyable. I check the site daily (via Pointcast News) and find some intriguing info, and to be honest, this was a field that was barely on the edge of my general interests. I have not been disappointed thus far. For David in particular: I find your reviews of the ASR software quite good, and useful. I only wish I had stumbled across your reviews earlier. I recently purchased IBM's VoiceType Simply Speaking Gold, for $99 American, and well, it has yet to prove its value to me. I knew it was not an ASR, but I was hoping to have some reliability with it. Strangely enough, I've found that the longer I worked with it, the WORSE it got. When I opened it up, and did an initial install, I immediately sent e-mails to my friends, using the software. They came out slightly garbled, but I'd say at least 75% okay. And the 25% was discernable what I intended to say. Then I attempted to use it for a more serious use, to read in some text written a few years ago, so that I could update it and place it in my new Word Processor. It drove me nuts. I could not get it to reliably enter even one whole sentence. This combined with having to pause between words makes me wonder if anyone could ever find the software useful. In fact, I found it better at navigating the OS than handling actual text representations. Which ostensibly makes sense, since the command set would be limited, and not open ended. I do await your future reviews on ASR software, especially, I am hoping to find something that can run on a P120 with 32 megs, as that is what my laptop runs on, and upgrading is not imminently possible for me at this stage. Keep up the fantastic job. Sincerely, Ed Greaves
Ed, thanks for your feedback. I think your experience with SimplySpeaking is worth sharing, especially as you have documented the difficulties so well.
One of the things I should have mentioned, and which you've now reminded me of, is the laptop issue. Most laptops, I believe, have built-in sound chipsets that may not be fully equivalent to the sound cards in desktops. Dragon lists only 6 or 7 "approved" sound cards for desktops, which suggests to me that laptop chipsets may present a problem. My next machine will be a laptop (my last one was, and I miss it!), and I will be seriously miffed if I can't run ASR software on it.
Ed replied: Interestingly enough, after I wrote that feedback, I went and installed SimplySpeaking on my laptop (Original install was on my desktop, but I like to have everything on both machines, since I never use the two at the same time.) And I think that it might have had a problem with the chipset. I haven't called for tech support yet. It appears to run, but I could not get it to perform the microphone setup, which is supposed to be required. I'm not sure why, and I haven't taken the time to find out.
I haven't really put it to the test, however, so if I find any changes in how well the system works, I'll let you know.
Me: Please do. I assume from your active involvement with ASR that you, like me, think that ASR is going to be a very significant change to the interface -- at least as momentous as the change from ASCII screens to windows, so every bit of experience with ASR is interesting and worth reporting.
Postscript for all: I'd like everyone to know that my continuing experience with Dragon NaturallySpeaking is becoming a drag, and I don't know why. Instead of getting better at recognizing my speech, it seems to be getting worse (which was Ed's experience with the IBM software). I've gotten into a routine of enunciating each word separately, which is slow and defeats the "continuous speech" attribute of the software, but it makes fewer mistakes that way and is less wearing on the temper. I've re-done several of the speech training files, to no avail.