We have mentioned that life may be a quality we attribute to some things, even more than it is a quality of the things themselves. This issue is perhaps most salient in the consideration of computer programs that emulate life processes. An artificial life program does not have to resem- ble any “real” living thing; it might be considered living by some set of criteria but seem, well, completely strange by earthly biostandards, by what we know from nature as it is on earth. Things might be able to
learn, to reproduce, to evolve in an adaptive way, without resembling earthling life much at all. Conway’s Game of Life opened the door for in- tense study of systems that exhibit lifelike behaviors.
And whatarelifelike behaviors? It is sometimes said that a defining feature of life is its resistance to entropy. Entropyis disorder, which al- ways increases in a system. Well-organized patterns fall into decay, disin- tegration, deterioration, as rust never sleeps. Yet life alone seems contin- ually to renew itself. Wolfram’s Type 4 CAs seem to have the ability to perpetually renew themselves, and it is for this reason that they have be- come the symbol, if not the seed, for artificial life research. They seem re- sistant to entropy.
It is possible to apply evolutionary algorithms to CA rules to produce behaviors of particular types, “breeding” interesting CAs. For instance, random mutations can be introduced in order to explore variations on CA rules by flipping a bit in the rule table occasionally—not too often, or you will destroy the behavior, but a low rate of mutation might allow the evolution of interesting dynamics. In this case we are considering the rule set as a kind of chromosome, a string of genes encoding the behavior of the cellular automata. This is a good analogy to the distinction be- tweengenotypeandphenotypein nature: you can’t see the genotype; it is an abstract coding scheme or program for the organism. What you see in nature is the phenotypic expression of the genotype. Just so in cellular automata, a researcher might encode the rule set in a string of ones and zeroes and mutate that string to evolve surprising behaviors in the visible cells.
One early program that could be called artificial life was produced as a sort of demonstration of the effects of genetic mutation on phenotypic forms. InThe Blind Watchmaker,Richard Dawkins (1987) created graphi- cal creatures he called biomorphs(see Figure 1.4), whose form was en- coded in a chromosome of nine genes, each of which could take on a value from zero to nine. These genes encoded the rules for the develop- ment of the biomorph, for instance, the angle or length of a branch, and they could be “mutated” by one step. That is, a gene in state 5 can, through mutation, be changed to a 4 or a 6. The user could generate a population of biomorphs with random genes, then select one that looked interesting. Dawkins reported that, though he expected his or- ganisms to look like trees, he discovered that insect forms were also pos- sible. Thus, he might select a biomorph that looked particularly insect- like. Once selected, this biomorph reproduces to create a generation of children, each containing a random mutation that causes it to differ slightly from its parent. The user can then select an interesting-looking
mutations resemblingit,and so forth. The result is incremental and un- predictable change in a desired direction: evolution.
In nature of course it is the environment that decides what individu- als will produce the next generation. Individual organisms with some new genetic pattern, that is, a mutation, may or may not reproduce, de- pending on the effect of the mutation in combination with the rest of the genetic heritage in the environment. In Dawkins’ program the re- searcher “plays God” with the population, deciding who lives and who dies, guiding evolution from step to step according to whatever whim or goal he or she has.
It is important to think about the size and structure of thesearch space for a population of biomorphs. With nine genes, each of which can exist in one of 10 states, we can see that there are 910, that is, 3,486,784,401, possible genetic combinations. These can be thought of as points in a nine-dimensional space. Though of course we can’t visualize such a space, we can visualize objects with attributes selected from locations on each of those nine dimensions. If mutation consists of making a change on a single dimension, and if the changes are (as in Dawkins’ program) simple moves to positions that are a single unit away on the dimension, then we can say that the child isadjacentto the parent. That means that
28 Chapter One—Models and Concepts of Life and Intelligence
conceptually, mathematically, they are near one another in the search space. This conceptualization turns out to be important when we want to search for a good point or region in the space, for instance, in search- ing for the solution to a problem. Whether the objects are continuous, binary, or—as in the current case—discrete, the concept of distance per- tains in searching the space; in particular, the concept applies when we want to decide what size steps to take in our search. If we can’t look at ev- ery point in a space, as in this case where there are billions of possibili- ties, then we must come up with a plan for moving around. This plan will usually have to specify how big steps should be in order to balance the need to explore or look at new regions, versus the need to exploit knowledge that has already been gained, by focusing on good areas. Dawkins’ biomorphs search by taking small steps through the space of possible mutations, so that the creatures evolve gradually, changing one aspect at a time. In nature as well, adaptive changes introduced by muta- tion are generally small and infrequent.
Other artificial life researchers have taken a slightly less godlike ap- proach to the evolution of computer-generated organisms. If the re- searcher provides information about the physics and other aspects of the environment, then life-forms can adapt to environmental demands and to one another. The artificial creatures developed by Karl Sims (e.g., 1994) are a beautiful example of this. Sims creates artificial three- dimensional worlds with physical properties such as gravity, friction, collision protection, and elasticity of objects so that they bounce realisti- cally off one another upon colliding. “Seed” creatures are introduced into this world and allowed to evolve bodies and nervous systems. In or- der to guide the evolution, the creatures are given a task to measure their fitness; for instance, in one version the goal is to control a cube in the center of the world. The creatures are built out of rectangular forms whose size, shape, number, and arrangement is determined by evolu- tion. Their “minds” are built out of artificial neurons composed of math- ematical operations that are also selected by evolution.
The randomly initialized creatures in Sims’ programs have been shown to evolve surprising strategies for controlling the cube, as well as innovative solutions to problems of locomotion on land and in water, with all forms of walking, hopping, crawling, rolling, and many kinds of fin waving, tail wagging, and body wriggling in a fluid environment. Some movies of these creatures are available on the web—they are well worth looking at.
It is possible then, in computer programs, to evolve lifelike beings with realistic solutions to the kinds of problems that real life-forms must
method for generating new variations, and some rules for capitalizing on adaptive improvements.
Christopher Langton (1988) has defined artificial life as “the study of man-made systems that exhibit behaviors characteristic of natural living systems.” In the same paper, he states, “Life is a property of form, not matter . . .” If we accept Langton’s premise, then we would have to admit that the similarity between an artificial-life program and life itself may be somewhat stronger than the usual kind of analogy. We will avoid declar- ing that computer programs live, but will note here and in other places throughout this book that it is unsettlingly difficult sometimes to draw the line between a phenomenon and a simulation of that phenomenon.