Darwin's Dangerous Idea (39 page)

only do what its programmer tells it to do." From our perspective we can see What would a good evaluation function look like? Some trivial games, like that this familiar but mistaken idea is nothing but an expression of Locke's tic-tac-toe, have feasible algorithmic solutions. There is a guaranteed win or hunch that only Minds can Design, an exploitation of
ex nihilo nihil fit
that draw for one player, and this best strategy can be computed in realistic Darwin had clearly discredited. The way Samuel's program transcended its amounts of time. Checkers is not such a game. Samuel (p. 72) pointed out creator, moreover, was by a strikingly classical process of Darwinian that the space of possible checkers games has on the order of 1040 choice-evolution.

points, "which, at 3 choices per millimicrosecond, would still take 1021

Samuel's legendary program is thus not only the progenitor of the intel-centuries to consider." Although today's computers are millions of times faster than the lumbering giants of Samuel's day, they still couldn't make a dent on the problem using the brute-force approach of exhaustive search. The search space is Vast, so the method of search must be "heuristic"—the 13. In fact, the bridge between computers and evolution goes back even farther, to branching tree of all possible moves has to be ruthlessly pruned by semi-Charles Babbage, whose 1834 conception of the "Difference Engine" is generally credited with inaugurating the prehistory of the computer. Babbage's notorious
Ninth Bridgewa-intelligent, myopic demons, leading to a risky, chance-ridden exploration of a
ter Treatise
(1838) exploited his theoretical model of a computing engine to offer a tiny subportion of the whole space.

mathematical proof that God had in effect programmed nature to generate the species!

Heuristic search is one of the foundational ideas of Artificial Intelligence.

"On Babbage's smart machine any sequence of numbers could be programmed to cut in, One might even define the task of the field of AI as the creation and inves-however long another series had been running. By analogy, God at Creation had appointed new sets of animals and plants to appear like clockwork throughout history—he had created the laws which produced them, rather than creating them direct" ( Desmond and Moore 1991, p. 213). Darwin knew Babbage and his
Treatise,
and even attended his parties in London. Desmond and Moore (pp. 212-18) offer some tantalizing glimpses 14. Samuel's 1959 paper was reprinted in the first anthology of Artificial Intelligence, into the traffic of ideas that may have crossed this bridge.

Feigenbaum and Feldman's classic,
Computers and Thought
( 1964 ). Although I had read More than a century later, another London society of like-minded thinkers, the Ratio that paper in Feigenbaum and Feldman when it first came out, I had, like most readers, Club, served as the hotbed for more recent ideas. Jonathan Miller drew my attention to passed over most of the details and savored the punch line: a 1962 match between the the Ratio Club, and urged me to research its history in the course of writing this book,

"adult" program and Robert Nealey, a checkers champion. Nealey was gracious in defeat: but I have not made much progress to date. I am tantalized, however, by the 1951

"In the matter of the end game, I have not had such competition from any human being photograph of its membership that graces the front of A. M. Uttley's
Information Trans-since 1954, when I lost my last game." It took a superb lecture by my colleague George
mission in the Nervous System
(1979 ): Alan Turing is seated on the lawn, along with the Smith in an introductory course in computer science that we cotaught at Tufts to rekin-neurobiologist Horace Barlow (a direct descendant of Darwin, by the way); standing dle my interest in the details of Samuel's article, in which I find something new and behind are Ross Ashby, Donald MacKay, and other major figures of the earliest days of valuable every time I reread it.

what has become cognitive science. It's a small world.

210 B

The Computer That Learned to Play Checkers
211

IOLOGY IS ENGINEERING

tigation of heuristic algorithms. But there is also a tradition within computer 1 for each string of passive pieces that occupy three adjacent diagonal science and mathematics of
contrasting
heuristic methods with algorithmic squares." At any one time, sixteen of the terms were thrown together into the methods: heuristic methods are risky, not guaranteed to yield results, whereas working genome of the active polynomial and the rest were idle. By a lot of algorithms come with a guarantee. How do we resolve this "contradiction"?

inspired guesswork and even more inspired tuning and tinkering, Samuel There is no contradiction at all. Heuristic algorithms are, like all algorithms, devised rules for elimination from the tournament, and found ways of mechanical procedures that are guaranteed to do what they do, but what they keeping the brew stirred up, so that the trial-and-error process was likely to do is engage in risky search! They are not guaranteed to
find
anything—or at hit upon good combinations of terms and coefficients and recognize them least they are not guaranteed to find the specific thing sought in the amount of when it did. The program was divided into Alpha, a rapidly mutating pioneer, time available. But, like well-run tournaments of skill, good heuristic and Beta, a conservative opponent that played the version that had won the algorithms
tend
to yield highly interesting, reliable results in reasonable most recent game. "Alpha generalizes on its experience after each move by amounts of time. They are risky, but the good ones are good risks indeed.

adjusting the coefficients in its evaluation polynomial and by replacing terms You can bet your life on them (Dennett 1989b). Failure to appreciate the fact which appear to be unimportant by new parameters drawn from the reserve that algorithms can be heuristic procedures has misled more than a few critics list" (Samuel 1964, p. 83).

of Artificial Intelligence, In particular, it has misled Roger Penrose, whose At the start an arbitrary selection of 16 terms was chosen and all terms views will be the topic of chapter 15.

were assigned equal weights__ During [the early rounds] a total of 29

Samuel saw that the Vast space of checkers could only be
feasibly
ex-different terms was discarded and replaced, the majority of these on two plored by a process that riskily pruned the search tree, but how do you go different occasions __ The quality of the play was extremely poor. During about constructing the pruning and choosing demons to do this job? What the next seven games there were at least eight changes made in the top readily programmable stop-looking-now rules or evaluation functions would listing involving five different terms __ Quality of play improved steadily have a better-than-chance power to grow a search tree in wise directions?

but the machine still played rather badly ___ Some fairly good amateur Samuel was searching for a good algorithmic searching method. He pro-players who played the machine during this period [after seven more ceeded empirically, beginning by devising ways of mechanizing whatever games] agreed that it was 'tricky but beatable'. [Samuel 1964, p. 89]

obvious rules of thumb he could think of. Look before you leap, of course, and learn from your mistakes, so the system should have a memory in which Samuel noted (p. 89) that, although the learning at this early stage was to store past experience. "Rote learning" carried the prototype quite far, by surprisingly fast, it was "quite erratic and none too stable." He was discov-simply storing thousands of positions it had already encountered and seen the ering that the problem space being explored was a rugged fitness landscape in fruits of. But rote learning can only take you so far; Samuel's program which a program using simple hill-climbing techniques tended to fall into confronted rapidly diminishing returns when it had stored in the neighbor-traps, instabilities, and obsessive loops from which the program could not hood of a million words of description of past experience and began to be recover without a helping nudge or two from its designer. He was able to overcome with indexing and retrieval problems. When higher or more recognize the "defects" in his system responsible for these instabilities and versatile performance is required, a different strategy of design has to kick in: patch them. The final system—the one that beat Nealey—was a Rube Gold-generalization.

berg amalgam of rote learning, kludges,15 and products of self-design that Instead of trying to find the search procedure himself, Samuel tried to get were quite inscrutable to Samuel himself.

the computer to find it. He wanted the computer to design its own evaluation function, a mathematical formula—a polynomial—that would yield a number, positive or negative, for every move it considered, such that, in 15. Pronounced to rhyme with "stooge," a kludge is an
ad hoc
or jury-rigged patch or general, the higher the number, the better the move. The polynomial was to software repair. Purists spell this slang word "kluge," drawing attention to its (likely) be concocted of lots of pieces, each contributing positively or negatively, etymology in the deliberate mispronunciation of the German word
klug(e),
meaning multiplied by one coefficient or another, and adjusted to various other

"clever"; but according to
The New Hacker's Dictionary
(Raymond 1993), the term may circumstances, but Samuel had no idea what sort of concoction would work have an earlier ancestor, deriving from the Kluge paper feeder, an "adjunct to mechanical well. He made some thirty-eight different chunks—"terms"—and threw them printing presses" in use as early as 1935. In its earlier use, it named "a complex and puzzling artifact with a trivial function." The mixture of esteem and contempt hackers into a "pool." Some of the terms were intuitively valuable, such as those exhibit for kluges ("How couid anything so dumb be so smart!") perfectly reproduces giving points for increased mobility or potential captures, but others were the attitude of biologists when they marvel at the perversely intricate solutions Mother more or less off the wall—such as "DYKE: the parameter is credited with Nature so often discovers.

212 BIOLOGY IS ENGINEERING

Artifact Hermeneutics, or Reverse Engineering
213

It is not surprising that Samuel's program caused a tremendous sensation, a function, sometimes they overlook retrospectively obvious shortcuts. Still, and greatly encouraged the early visionaries of AI, but the enthusiasm for optimality must be the default assumption; if the reverse engineers can't such learning algorithms soon faded. The more people looked into the assume that there is a good rationale for the features they observe, they can't attempt to extend his methods to more complex problems—chess, for even begin their analysis.16

instance, to say nothing of real-world, non-toy problems—the more the Darwin's revolution does not discard the idea of reverse engineering but, success of Samuel's Darwinian learner seemed to be attributable to the rather, permits it to be reformulated. Instead of trying to figure out what God relative simplicity of checkers rather than to the power of the underlying intended, we try to figure out what reason, if any, "Mother Nature"— the learning capacity. Was this, then, the end of Darwinian AI? Of course not. It process of evolution by natural selection itself—"discerned" or "dis-just had to hibernate for a while until computers and computer scientists criminated" for doing things one way rather than another. Some biologists could advance a few more levels of complexity.

and philosophers are very uncomfortable with any such talk about Mother Today, the offspring of Samuel's program are multiplying so fast that at Nature's reasons. They think it is a step backwards, an unmotivated conces-least three new journals have been founded in the last year or two to provide a sion to pre-Darwinian habits of thought, at best a treacherous metaphor. So forum:
Evolutionary Computation, Artificial Life,
and
Adaptive Behavior.
The they are inclined to agree with the recent critic of Darwinism, Tom Bethell, in first of these emphasizes traditional engineering concerns: using simulated thinking there is something fishy about this double standard (see page 73 ). I evolution as a method to expand the practical design powers of programmers claim that it is not just well motivated; it is extremely fruitful and, in fact, or software engineers. The "genetic algorithms" devised by John Holland unavoidable. As we have already seen, even at the molecular level you just (who worked with Art Samuel at IBM on his checkers program) have can't do biology without doing reverse engineering, and you can't do reverse demonstrated their power in the no-nonsense world of software development engineering without asking what reasons there are for whatever it is you are and have mutated into a phylum of algorithmic variations. The other two studying. You have to ask "why" questions. Darwin didn't show us that we journals concentrate on more biologically flavored research, in which the don't have to ask them; he showed us how to answer them (Kitcher 1985a).

simulations of evolutionary processes permit us, really for the first time, to Since the next chapter will be devoted to defending this claim by dem-study the biological design process itself by
manipulating
it—or, rather, by onstrating the ways in which the process of evolution by natural selection is manipulating a large-scale simulation of it. As Holland has said, Artificial Life
like
a clever engineer, it is important that we first establish two important programs
do
permit us to "rewind the tape of life" and replay it, again and ways in which it is
not
like a clever engineer.

again, in many variations.

When we human beings design a new machine, we usually start with a 6. ARTIFACT HERMENEUTICS, OR REVERSE ENGINEERING

16. This fact has been exploited by counter-reverse-engineers. I discuss an example in The strategy of interpreting organisms as if they were artifacts has a lot in Dennett 1978 (p. 279):