Excerpt from my book Machine Nature
Tyger! Tyger! burning bright
In the forests of the night,
What immortal hand or eye
Could frame thy fearful symmetry?
Who indeed framed the tyger? Two hundred years after William Blake wrote the beautiful opening stanza of The Tyger, we have a better idea of how tigers come about: through the process of evolution. There is neither a Master Plan, nor a "hand of god," nor any ultimate goal; the driving force is the short-term objective of survival, with the process consisting of the slow accumulation over millenia of numerous small — yet profitable — variations. In the forests of evolution burn many a creature, with nature's immortal evolutionary hand slowly framing the fearful symmetry of the tiger.
Natural evolution is an open-ended process, and is thus distinguished from artificial evolution which is guided, admitting a "hand of god": the (human) user who defines the problem to be solved. When we apply evolution with a well-defined goal in mind — such as designing a bridge, constructing a robotic brain, or developing a computer program — what we are doing is akin to animal husbandry. Farmers have been using the power of evolution for hundreds of years, in effect doing evolutionary computation on domestic animals. In order to "design," say, a faster horse they mate swift stallions with speedy mares, seeking to see even faster offspring emerge from this coupling. This is quite similar to the use of evolutionary techniques discussed in this book: the farmer defines the fitness criterion (say, speed) and performs the selection process by hand (by choosing the fastest individuals in the equine population); he then lets nature work out the genetic details involved in the coupling act. It's rather interesting to note that farmers and breeders had started using this method long before either evolution or genetics came under the scrutiny of science.
Farmers can start out with slow horses and evolve fast ones — but can they evolve tigers? Engineers can start out with bad bridges and evolve good ones — but can they evolve a town like Cambridge? Robotics researchers can evolve robots that manage to amble decently — but can they evolve a robotic housemaid. Programmers can evolve computer programs that solve various well-defined problems — but can they evolve truly intelligent software? Natural evolution has done it all: complex organisms, sophisticated structures, intelligent beings; and it did so by being open-minded, ready to accommodate any improvement that came along.
The Merriam-Webster online dictionary (www.m-w.com) defines "open-ended" as something that is "not rigorously fixed: as a : adaptable to the developing needs of a situation b: permitting or designed to permit spontaneous and unguided responses." Open-endedness is thus the flip side of guidedness, and it is a crucial aspect of natural evolution. Since nature has no specific goal in mind she can easily change course so as to face the winds of change, and in so doing she explores numerous designs out of what is essentially an infinitude of possibilities. Man, on the other hand, even when using evolutionary techniques does have an ultimate goal in mind — be it a retractable bridge or a program that computes taxes.
When we apply evolutionary techniques the ingredients are all there: a (possibly huge) population of individuals, survival of the fittest, and the equivalent of genetic operators. Yet the hand of god is ever-present in the background: at every step of the way an individual's fate is decided in accordance with its ability to perform in the arena set up by the puppet master; and the master wants his puppets to do some very specific tricks. This places a fundamental a priori limit on what evolution can achieve: if we set about to find fast horses, then we might succeed in doing so — but we'll not suddenly see the emergence of tigers.
Nature's open-endedness runs deeper, though, than the mere absence of a goal and a god — of a teleology and a master. It has to do with her ability not only to play the game but indeed to change the rules altogether. Let me drive this point home by way of a sportive example. The game of basketball is played on a court 90 feet long by 50 feet wide between two opposing teams of five players, who score by tossing an inflated ball through a raised goal. The rules are well known and rigid, with changes being rare, minor (for example, adding the three-point shot), and human-mediated (say, the NBA committee). This scenario is analogous to that of guided evolution: a human designer sets the stage (or in this case court) that gives rise to a (fiercely competitive) evolutionary process, from which but one kind of creature may emerge: basketball players. The process is not open-ended since there is a precisely defined goal (scoring more points), with virtually immutable rules. Though superb basketball players can (and do) evolve, this arena does not give rise to first-rate opera singers.
Playing nature's "basketball" game is quite different. For one thing, there is no clear objective; at best, one can speak of a very basic goal, that of coming out of the match alive. What's more — and this is where open-endedness comes into play — nature keeps changing the rules of the game, both in time and in space. Being seven foot tall might be good at a certain place and time, whereas elsewhere or else-when it might be downright deleterious. And sometimes the rules are such that having a superb tenor voice is a match winner. Nature's game of basketball is more of a meta-game, where you want to score more points — but have to figure out how points are scored.
In Chapter 1 we discussed an important distinction in nature, that between genotype and phenotype. An organism's genotype is its genetic constitution, the DNA chain that contains the instructions necessary for the making of the individual. The phenotype is the mature organism that emerges through execution of the instructions written in the genotype. It is the phenotype that engages in the battle for survival, whereas it is the genotype — safely cached in each cell of the organism — that accrues the evolutionary benefits.
Setting a specific goal — as with artificial, guided evolution — means that there is a highly restricted environment; the basketball-player phenotype faces an environment in which it is demanded to perform a very specific task: playing basketball. Natural environments are not only much more complex but also highly dynamic — the phenotypes must face ever-changing circumstances.
Nature's open-endedness manifests itself not only at the phenotypic level but also at the genotypic level: not only can the rules of the playground change, but so too can the rules for making players. The genome of a red ant is quite different from that of an orangutan (though as both are branches of the Tree of Life, they also bear many similarities). As we've seen, artificial-evolution scenarios to date are limited, being goal-oriented, with but very little maneuverability in changing the genetic makeup. A bridge genome will always produce a bridge — perhaps a superb one at that — but never a skyscraper. Nature, though, can tinker with the genome, thus changing the underlying construction plan, so as to produce entirely different beings, including skyscrapers (giraffes) and towns (ant colonies). This is a crucial aspect of her open-endedness.
We've seen how artificial evolution is used to design complex objects, which stretch — or overstretch — our classical engineering techniques. The results are often quite impressive and at times those who use them are even reputed to cry out: "Wow, I'd have never come up with such a solution." But this is still at the level of evolving super bridges or superb basketball players; moreover, it might even be limited at that since an entirely novel bridge design or a new form of basketball player might necessitate genomic tinkering that is beyond the system's reach. Can something truly astounding — something entirely new — emerge out of an artificially set stage? In my mind this is one of our grandest challenges, and it may still be many years in the coming. I like to think of this defy as that of building a system that Knocks Your Socks OFF. Following the time-honored tradition in computing science of coining acronyms, this might be dubbed the KYS OFF challenge — which leads me to wonder whether such a system would kiss us off ...
As our artifacts become more and more complex, so does their design become more arduous. One way out is to employ the powerful process of open-ended evolution. But wait a minute — by definition, that would mean ... removing the designer from the equation! Then who controls the design process — who's the boss? It seems that you can't have your cake and eat it too — something has to give. With guided evolution the guide — or designer — maintains a great deal of control over his system, and though he'll often be overwhelmed by the results obtained, his socks will remain firmly in place. Open-ended evolution might indeed knock your socks off, but at the price of giving up some of that precious control that we've grown used to.
Strangely enough then, it is less design, meaning more open-endedness, that increases our design power. Uhm ... did I just say less design? Actually, you have to set up the stage so as to be more open-ended, that is, you have to design the system to exhibit ... less design! That's the essence of the KYS OFF challenge, which only nature has met so far — but then again, she's been at it for the past threenand a half billion years. (While I've been concentrating my discussion of the open-ended versus the guided on evolution this is by no means the only process of interest. Learning, for one, augments a system's open-endedness.)
Open-ended goes hand in hand with less control — though with the potential of more spectacular results. Parents usually want their children to grow up to be independent and able to think for themselves. But in many ways child-raising is open-ended — with no guarantees: What if the child decides to be a rock star? (Result: horrified parents.) Or a doctor? (Result: delighted parents.) In Chapter 4 we discussed the application of biological processes, such as evolution and learning, in the field of adaptive robotics. We saw that one of the central goals is that of attaining more autonomous robots; I doubt, however, that we're ready to see them declare autonomy ...
With an open-ended process not only do you not control the precise shape that the final outcome will take, you're not even sure what this outcome will be. When we look at nature's magnificent products with awe and with envy, we should always keep in mind the billions of years that their production necessitated. If you set off such a process and then patiently wait for a couple of billion years, you might find — a posteriori — lots of wonderful devices, such as eyes, toes, flowers, brains, and wings. You might be quite happy with this plethora of gadgets that will bring you fame and fortune. But this process is open-ended, which means you don't know in advance what the final products will be. In fact, it might not even get off the ground: it took nature almost three billion years before things really started to pick up and the Tree of Life began to grow. What if it never gets off the ground? Or if you simply get tired of waiting?
The possibility of creating an artificial scenario in which open-ended evolution takes place is at the heart of Greg Egan's excellent science fiction novel Permutation City. Explaining to the researcher her mission, the protagonist says: "I want you to construct a seed for a biosphere ... I want you to design a pre-biotic environment — a planetary surface, if you'd like to think of it that way — and one simple organism which you believe would be capable, in time, of evolving into a multitude of species and filling all the potential ecological niches.'' Having succeeded in creating such a biospheric seed, evolution is then set lose in this artificial universe known as the Autoverse, to work its magic over the eons: "We've given the Autoverse a lot of resources; seven thousand years, for most of us, has been about three billion for Planet Lambert.'' And the outcome? "There are six hundred and ninety million species currently living on Planet Lambert. All obeying the laws of the Autoverse. All demonstrably descended from a single organism which lived three billion years ago — and whose characteristics I expect you know by heart. Do you honestly believe that anyone could have designed all that? " The answer is no; be it in an artificial or a natural world, open-ended evolution will knock your socks off ...
If we give up our control, can't things get out of hand, leading to a system run amok? This is a tough question, which needs to be addressed on a case-per-case basis. We should come up with fail-safes (à la Asimov's three laws of robotics). We might well wish to place checks and bounds (for example, limit the robots' autonomy). We'd like to maintain the possibility of pulling the plug if things get downright ugly. But this issue is by no means an open-and-shut case: how does one juggle between control and autonomy — between guided and open-ended? This issue will probably gain more prominence as our technology advances, enabling us to build systems that are somewhat less controlled — and less controllable.
Which brings me back to William Blake and the closing lines of "The Tyger,'' where the "Could'' of the first stanza has been conspicuously replaced by "Dare":
Tyger! Tyger! burning bright
In the forests of the night,
What immortal hand or eye
Dare frame thy fearful symmetry?
Dare we frame a tyger?