Touched by nature
Putting evolution to work on the assembly line

BY CHARLES W. PETIT

A new kind of evolution is on the loose, and to hear its practitioners talk, the prospects are surreal. "Mom and dad jet engine can get together and have baby jet engines. You find the ones that work better, mate them, and just keep going," says David Goldberg, a professor of engineering at the University of Illinois. He is a leader among researchers who, with little fanfare, have hijacked evolution from the world of the living. Stripped down and souped up, this new evolution is ready, after 30 years of gestation, to go to work as an industrial, invention-spewing tool.

Evolution as in Charles Darwin, blind chance, survival of the fittest, and all that? Yes. This is the same descent-with-modification evolution, right down to the lingo--sex, parents, offspring, selection, mutations, genes, and chromosomes--that biologists use to explain the emergence of new species. Except in this case, the product is not living tissue but complex hardware, solutions to maddeningly difficult scheduling problems, or novel molecules that evolve out of computer code, or even DNA.

Breeding turbines. The Boeing Co.'s 777 airliner has a General Electric engine whose turbine geometry evolved inside a computer, and the company is experimenting with evolving wings for future airliners. Eli Lilly and other pharmaceutical companies use "directed evolution" to find new protein catalysts to help produce drugs faster; Deere & Co. breeds daily schedules that direct assembly lines in six factories to fill custom orders for its millions of variants of agricultural machinery. The government contracted with Natural Selection Inc. of La Jolla, Calif., to use evolutionary programming in computers that will read mammograms more quickly and inexpensively than a radiologist.

Applying biological principles to engineering isn't as tough as it sounds, but it requires computing muscle that has only recently been available. About five years ago, Andrew Keane, a professor of engineering at the University of Southampton in England, took a hard look at a prototype space-station girder assembled by American astronauts aboard the space shuttle in 1985. Keane had read Goldberg's work on computer-based evolution. Much of modern engineering uses algorithms--mathematical procedures for solving problems. But Goldberg is a champion of genetic algorithms, which use computers to manipulate potential solutions as if they were living organisms. Keane wondered if genetic algorithms could outdo NASA's human engineers. To find out, he recast the original design of the girder as strings of numbers describing thickness, angle of attachment, and other aspects. He called each number a gene, each string of numbers a chromosome--analogues to the DNA genes and chromosomes that orchestrate living cells. Keane then copied his digital truss "genome" enough times to produce a diverse founding population. Finally he said, in effect, "Let there be life," and ran the program on 11 interconnected computer workstations. For several days, the truss designs had cybersex--they swapped digital genes with random abandon. To be sure, Keane, creator of this pseudo world, imposed his influence over the breeding. He had defined ahead of time what constituted fitness, and the computers tested each emerging design accordingly. Those that suppressed vibration best yet remained lightweight and strong were rewarded with greater fertility. Generation by generation, the fittest got fitter. The program threw occasional random mutations among the competing genomes to provide a little extra variety.

Thus there emerged, from 15 generations and 4,500 different designs, a truss no human engineer would design. The lumpy, knob-ended assembly reminds Keane of a leg bone, irregular and somehow organic. Tests on models confirm its superiority to human-designed ones as a stable support. No intelligence made the designs. They just evolved.

Impressed by Keane's work, executives of Matra Marconi Space, a French-British satellite manufacturer, last year signed his group to help design an orbiting infrared telescope platform for finding planets similar to Earth around other stars. Its name: Darwin. "It's a remarkable irony that we may wind up looking for life on other planets, using mechanisms made by the process that created life here. It sends tingles up my spine," Keane says.

Engineers expect that similar techniques will reap a bonanza of innovations here on Earth. "By the middle of the next century, there will be no area of engineering not touched" by these genetic methods, says J. David Schaffer, a senior researcher at the Philips Electronics company's North American research center in Briarcliff Manor, N.Y. "The only way to get to the next level of complexity is with evolutionary methods."

No single term has arisen to label this new kind of evolution. Variants of it include not only genetic algorithms but directed evolution, evolutionary programming, evolution strategy, and evolutionary computation. In Madison, Wis., this week, more than 400 specialists will turn out for the Third Annual Genetic Programming Conference to discuss the field's growing success in industry.

The code of life. Oxford University evolutionary biologist Richard Dawkins saw the border between life and machine start to blur more than 10 years ago. In his 1986 book, The Blind Watchmaker, Dawkins wrote: "What lies at the heart of every living thing is not a fire, not warm breath, not a 'spark of life.' It is information, words, instructions. There is very little difference, in principle, between a two-state binary information technology, like ours, and a four-state information technology like that of the living cell."

Now, engineers who formerly thought in Cartesian, gears-and-straight-line terms are finding the blending of biology with hardware and software to be liberating. Bill Fulkerson of Deere & Co., who shepherded his company's use of genetic algorithms, says, "The old metaphors were mechanical, shoulder to the wheel, and all that. When you open up to biology, and to new metaphors like ecologies of companies, you get a completely different perspective on how things work." At Deere, factory-floor supervisors key into ordinary PCs the list of hay balers, air-conditioned tractors, and other customized farm machinery on order, and the software sets a swarm of prototype schedules loose. In a few hours, a list emerges deciding which machines to make when, a list consistently more efficient than any person could have figured out.

Those sorts of vexing practical problems are driving much of the work in evolutionary computation. Goldberg, the University of Illinois engineer whose 1989 textbook is the bible of the field, got his start with the topic out of frustration with his job consulting for a natural-gas-transport company in the 1970s. Efficient management of intricate pipeline networks seemed impossible. Hoping that artificial intelligence could help, Goldberg went back to school and wound up in the classroom of John Holland at the University of Michigan. The first lectures were all biology, with nothing, it seemed, to do with engineering. Then it hit Goldberg: "Maybe this biology stuff has everything to do with everything." Holland, still a central figure in evolutionary computation, invented genetic algorithms in the early 1960s, but they and similar methods languished until the late 1980s.

With the advent of greater computer power, Holland's principles now are being applied broadly, even to accelerate research with actual raw genes, the stuff of "real" evolution. "I call it Darwin in a test tube," said Frances Arnold, a California Institute of Technology professor of chemical engineering. For 10 years she has been developing directed evolution, which scrambles real snippets of DNA, mutates some of them, crosses them with one another in a process like sex without the good parts, then plants them back in microbes and harvests new proteins. A relatively small protein containing 300 amino acids can have far more variants than there are protons in the universe. A human's attempt to design a new protein, Arnold says, "is fruitless, doomed to failure because our puny brains cannot understand the systems we want to design."

But in just three cycles of DNA shuffling, researchers at Maxygen Inc. of Santa Clara, Calif., using a method similar to Arnold's, applied directed evolution to a protein for antibiotic resistance in bacteria. The result was a version that worked 32,000 times better than the protein present in microbes naturally. Milton Zmijewski, a senior research scientist at Lilly Research Laboratories' drug labs in Indianapolis, said directed evolution is perfect for his company. "We don't care how we get there, as long as we get there first and fast."

Computer ooze. Some of the new evolution even uses computer programs to breed their own progeny, swapping software code like genes. South of San Francisco is the hilltop aerie of John Koza, a Stanford computer scientist who, as cofounder of Scientific Games Inc., made a fortune by inventing the scratch-off lottery ticket. Now he has loaded a room in his huge split-level home with 70 networked computer processors, each running at 533 megahertz, or half a billion calculations per second, and expects to have 1,000 processors by year's end. He is loading them with evolvable computer programs for industrial use, including ones that will design electronic circuits and control robots. Koza calls his method "genetic programming," a version of genetic algorithms. While the latter use computer programs to manipulate strings of numbers representing real-world things, Koza's technique allows the programs themselves to crossbreed and evolve. His start-up programs are random snippets of code, what Koza calls primordial computer ooze. As ensuing generations become increasingly effective, they also begin looking more bizarre. They accumulate seeming garbage, circles of illogic that no human would think up--just as the DNA of living things is thronged with stretches of nonsense and leftovers of forgotten ancestors.

According to Koza, it is precisely because evolutionary computer code is messy that it finds solutions that are more subtle and flexible than does that written by a human programmer. "In nature, nothing is brittle; it is smooth and elegant, because you never encounter exactly the same situation twice," Koza says.

To be sure, there remains a vast gulf between the subtlety of living creatures and the innards of even the most advanced computers or machines or individual proteins. No computer is even close to being able to cope with the interactions among the 100,000 genes in a human being. And it is hard work to frame even fairly simple problems so they can be genetically manipulated. Breeding and testing solutions can gobble hours to days of computer time. The human brain may always be better for solving some problems, but as computers get faster and faster, it is inevitable they will become breeding grounds for a growing share of invention.

The implications are profound, not only for engineering but for our view of ourselves. "As evolution becomes more a standard part of engineering techniques, and more and more people gain firsthand experience with an evolutionary process, people will feel more comfortable with the idea of evolution as the core of their own history," said Lee Altenberg, a research affiliate at the University of Hawaii.

But can one trust inventions that invent themselves, with people as mere interested observers? Philips's Schaffer recalls a meeting about genetic algorithms a few years ago. "Is anybody concerned that you might be living downwind from a nuclear power plant controlled by a robot that evolved?" Schaffer asked. Nobody, he recalls, was bothered. As one said, why worry? After all, today we turn such plants over to people. Nobody really knows how they work, either.