Darwin’s Devices

Chapter 2 THE GAME OF LIFE

“GREAT IS THE POWER OF STEADY MISINTERPRETATION.” This lament by Charles Darwin, from his sixth and final edition of The Origin of Species in 1872, summed up years of simmering frustration. Many of his critics and even some of his well-meaning champions had oversimplified his particular theory (other theories existed at the time) of descent with modification, what we now call evolution. The oversimplification was this: descent with modification has a single cause, natural selection.

Although natural selection was Darwin’s most important insight, he recognized and stated repeatedly in print that while it was the primary mechanism of change, it was not the only one. “Evolution by natural selection” was the phrase that I used in the previous chapter to define “adaptation.” Though that may be true, it’s only part of the evolutionary picture. Oops. Because I didn’t talk about other kinds of causal mechanisms—like mutation, recombination, genetic drift, and assortative mating—I’m one of the oversimplifiers. Let me make amends here to get you ready for the lifelike complexities of evolving robots.

I think that Darwin, a keen observer, would’ve loved watching our evolving robots. With them we can show what evolution looks like when selection is dominant, on the one hand, and when it takes a backseat to other evolutionary mechanisms, on the other. We can use robots to look at evolutionary processes, those ongoing, real-time, cause-and-effect interactions of autonomous agents with their environment—at any specific place and time. That is, we can become spectators at the greatest game on Earth: the game of life.

Think of it this way: life is a game, a never-ending contest played on the world’s stage. But the players are not often locked in open combat. Although a great white shark hunting a California sea lion makes for dramatic theater on Animal Planet’s “Shark Week,” in the real game of life most of the players never meet. Instead, each player is more like a plodding decathlete, doing ten different sports in quick succession and often at the same time. Each mobile, autonomous animal navigates its landscape, finds food or hunts for it, figures out how or if to eat what it’s found, detects and escapes threats, seeks and selects mates, finds shelter if it can, and makes offspring. Winners are those who survive long enough to reproduce. Among the winners, the champions are those who have the most children. The game of life is called evolution, and robots are allowed to play.

Evolution is one of the simplest games on the planet. It has only three rules:

* You score points for each child you create.

* You score bonus points if your children make offspring of their own.

* You can use any means to make children and to help your children make children.

Just because the rules are simple doesn’t mean that the strategies are simple too. Some players may realize that cooperation provides more eyes, ears, and noses for collecting food and avoiding predators.[1] Other players may figure out that if they don’t have to raise their offspring, they can spend much more time making them. Some may discover that, through deceit or cuckoldry, they can have others raise their children. Others, still, might focus on collecting and protecting the resources that they need to raise their children. Players may also figure out that selecting the right mate can make the difference between success and failure.

In the real game of life most players, even humans, don’t evaluate the evolutionary effectiveness of their behavior. How can we, when most of us don’t even know that we’re playing the game? Instead, the game is played through instinct, gut-level emotional reactions to the circumstances that present themselves. Even if you know that you’re a player, know the rules, and know what it takes to win, you can only guess as to what might make your children successful at making grandchildren. You can’t see the future and the chance events that may alter the conditions on the playing field. We are playing blind.

Although you as an individual may not score points through the three rules described above, you are, in fact, part of a team, and you get some points for just being part of a team that has players that are reproducing.[2] You can help other members of your familial team reproduce or raise children. But please don’t misunderstand: if you are personalizing this, thinking about your own play in the game of life, then you may be getting bummed out right about now, because I may have conveyed that without children and grandchildren you are a loser. That’s not my intent. I’m just trying to show you a different way to think about evolution—by thinking of it as “differential reproduction.”

Understanding the fundamentals of evolution is key because we so often get it wrong. Think of Darwin’s lament. Evolution is part of our everyday parlance, and even though the game of life is a fact of life, we intentionally and unintentionally misrepresent, steadily. We think, incorrectly, that individuals evolve, that individuals act for the good of the species, that some species are primitive and others are advanced, that a ladder of life describes descent with modification, and that evolution is always working to make species better. We incorrectly intuit that complexity is always more evolutionarily advanced than simplicity, that evolution is goal driven, that evolutionary change is linear and in one direction, that any anatomical structure evolved long ago for the function it fulfills today, and that humans aren’t evolving anymore—wrong, wrong, wrong!

These same poison apples will tempt us when we talk about evolving robots. Our intuitive, mistaken expectations will bear fruit in the form of disappointment, disapproval, and denial. So keep this in mind: nowhere in the rules of the game of life does it say that the players in the future will always be smarter or better in any way than those playing the game right now. The rules don’t say that different strategies that work at one time and place will work at other times and places. And the rules are silent regarding the behavior of other players toward one another, the number of players, the kinds of players, the availability of resources, the use of the environment, and accidents and other chance events that may occur. When robots evolve, we simply don’t know what will happen. That’s life.[3]

I have more bad news: individuals don’t evolve. As much as it makes good science fiction when Captain Jean-Luc Picard “devolves” into a lemur on Star Trek: The Next Generation, individuals are trapped in time and space. An individual human carries a genome—the total complement of genes and DNA within that individual—that is the product of genes from mom’s egg and dad’s sperm. This new combination of genes interacts with the world to create an embryo, just like our autonomous agents interact with the world to create behavior. The behavior of the genome, in this case, creates the ongoing interaction with the world known as development, the splitting of one cell into two, two cells into four, and so on, creating a multicellular animal from a single cell in a matter of a few hours.

The inside of each cell—with its aquatic world of chemicals in solution, lattice-like network of tiny skeletal structures, and membrane-bound micromachines—is the world of the genome. The genes interact with proteins that wind up and organize the DNA; the genes interact with other kinds of proteins that signal when the gene should make RNA; the DNA interacts with itself in order to make copies prior to cell division.

Each cell also has a world with which to interact. Other cells cling to it, pull on it, and exchange chemicals and electric charge. Fluid not in cells can be present in some tissues, and that extracellular fluid can bring hormones that the cell reacts to, setting up a cascade of molecular signals that results in changing how the genome is working. In response to being in different positions in the multicellular embryo, some cells “differentiate”—that is, their genes start making different kinds of RNA, and the RNA starts making different kinds of proteins. Those different proteins self-assemble into different structures. Quickly, you have cells in a particular neighborhood of the embryo that are working together to make a notochord, the skeletal rod that runs from head to tail in all vertebrate embryos and is retained in the adults of some fish and amphibian species. The cells making a notochord, in turn, release compounds that cause neighboring cells to start making a central nervous system and so on, throughout the lifetime of that individual: the genome, copied and partitioned into cells, interacts with its local world inside the cell, outside the cell, and embedded in differentiating tissues that are, in turn, interacting with the world outside the individual.

How, given all this developmental interaction of the genome and the world, is an individual trapped in time and space? Each individual is literally a product of their time and place (where time and place = “world” as I’m using the word). Take the same genome and put it in a different time and place, and you will get a different individual. Interaction of the genome and the world unfolds in development, and development reflects that particular history of that agent. Each agent is “trapped” in the sense that their agent-world interactions—which are unique—make them what they are as they continuously become what they are.

Sounds a lot like a self-help manual, eh? That self-help, you-can-change, Zen-transformation approach to our own individual histories leads us to equate our own development with evolution. That’s the crux of the problem for our intuition. While both development and evolution are change-over-time phenomena, what changes in each process is different. Allow me to oversimplify: in development it’s not the genome that’s changing but rather what the genome makes, the material substances of the individual. In evolution it’s the genome that changes, and in spite of the fact that we can have cell-level mutations that make different copies of the genome within an individual, the only way for the changes in the genome to have an evolutionary impact is for those changes to occur in egg or sperm and be passed on to the next generation.

Here’s what natural selection looks like. Individuals in a population coexist in time and place. Individuals differ in their anatomy, physiology, and behavior. Openly or unknowingly, individuals cooperate and compete with each other for sex, sustenance, safety, and shelter. Some individuals are better than others at cooperating and competing. Differences in anatomy, physiology, and behavior cause some of the differences in cooperation and competition. Thus, differences in anatomy, physiology, and behavior endow some individuals with an advantage over others in the game of life and the struggle for existence. Those advantageous differences enhance some individuals’ ability to survive and make offspring. That’s natural selection.

If those advantageous differences can be passed on to offspring—meaning that those differences are encoded in part by genes—then the next generation will look, function, and behave differently from the previous generation. That’s evolution by natural selection.[4]

As you can see, with natural selection, as we’ve just defined it, some individuals wind up having more offspring than other individuals do (look back at Figure 2.1). Individuals that out-reproduce others are said to have been “selected for,” and those that lose in the game of life are said to have been “selected against.” In this sense, every single individual of the same species, together at the same time and place—a grouping that biologists call a “population”—is “under selection” if the conditions described above hold.

So the bottom line is this: individuals can be selected, if the conditions are right, but they don’t evolve. This selection means that only some of the parents reproduce, so each successive generation looks as a group very different from the parental population as a group. The group, the population moving through generational time, is the entity that evolves. Sorry, you rugged individuals, but that’s the way the game of life is played.

Because you, as an individual, don’t evolve, passing on your genome—making babies—is the best you can hope to do in the evolutionary game of life. Individual life-forms can make babies in two ways. They can make multiple copies of themselves that have nearly identical genomes, a process that biologists call cloning, or asexual reproduction. Individual life-forms may take a second path, sexual reproduction, in which the individual produces eggs or sperm, known collectively as gametes, and engages in some process to put their gamete in close proximity to another gamete from the same kind of life-form. Most plants and animals reproduce sexually. Plants do this, as your parents told you, with flowers and pollen, sometimes with an animal, like a bee, acting as the intermediary. Animals reproduce sexually either by spawning or by depositing gametes in their partners.

Generally speaking, sexual reproduction brings together the genomes from two different individuals into one new individual; it is thought to be better than asexual reproduction in producing offspring that are variable (although some plants can fertilize themselves). Both asexual and sexual reproducers can also have mutations—changes in the genetic code—that can be passed on. To be passed on, the mutations have to occur in the cells that will make the offspring. For sexual reproducers that means mutations have to occur in the cells that create gametes. The making of gametes, a process known as gametogenesis, has several important features. One is that each gamete gets only half the parent’s genetic material, one from every pair of chromosomes (in humans, most cells have twenty-three pairs of homologous chromosomes, for a total of forty-six, and gametes have just twenty-three unpaired chromosomes). Another is that during gametogenesis a process known as crossing over, or recombination, occurs; essentially this means that the genetic material from one chromosome in a pair is shuffled to the other before the pair is split up and delivered to separate gametes. The result is both new (mutated and/or recombined) genes, and new combinations of genes in every gamete produced.

Sexual reproduction’s secret weapon is the final twist: bringing together sperm and egg. When sperm and egg meet, they create a single cell, called a zygote, which has half the genome of each parent. You can see right away why offspring from sexually reproducing parents are different and why sexual reproduction is such an excellent means of producing the variable populations required for evolution by natural selection to happen.

You’ve got enough information now to figure out how you can detect evolution in action. Think about measurement. What could you measure? Keep in mind that you’ve got to measure features of the population. You need to sample individuals and claim, usually with statistical reasoning, that the individuals you sampled represent the whole population. Or better yet, measure every individual in the population, as Rosemary and Peter Grant have done with ground finches on the island of Daphne Major in the Galapagos.

If you head out to Daphne Major with the Grants, you’ll see that they net finches, weigh them, and quickly measure the size and shape of their bodies with a pair of calipers, which is basically a high-resolution ruler.[5] They tag each individual with colored bands so that they can keep track of them. They spend hours and days observing males and females nesting together as the birds select and process food, lay eggs, and feed chicks. They measure and tag the chicks. The mountains of data, collected over years, are then analyzed for things like the average length of the bill in that generation of birds. The Grants can then look at how the average length of bills (and many other features) changes from generation to generation. They can also measure how the variability, what statisticians call variance, of the length of the bill changes over generational time.

When the average and/or variance of the length and thickness of the bill changes from one generation to the next, it is the first clue that natural selection and other evolutionary forces are at work in this particular population at this particular time and place. These visible physical and behavioral features of the birds are what biologists call “phenotypes.” Any phenotype may or may not have a genetic basis. If bill length has, at least in part, a genetic basis, then the change in the average bill length over generational time is evolution. The change in the average and variance of a phenotype within a population is one way to measure evolutionary change.

You can see, though, that we can run into trouble if we forget about our the conditions for evolution by natural selection. What if the phenotype doesn’t have a genetic basis? What if individuals learn some new trick that isn’t genetic? We can measure changes in the presence of the trick from generation to generation, so we think we are measuring evolutionary change, but upon closer inspection we find that the transmission of the behavior occurs by parents teaching their young how to do it. Orcas, for example, teach members of their pod how to specialize in hunting. Members of some pods eat otters. Members of other pods eat seals. Because we can observe the old teaching the young, we know that the tricks of the trade are learned rather than inherited genetically.

One way out of this problem is to focus, as many evolutionary biologists do, on the genotype. If the genes present in a population change, then we know that evolution has happened. We would do this by focusing on alleles. An allele is any particular version of a gene. If you have a gene that produces a protein, two different alleles of the gene may cause the protein to have a different shape or other properties. Every allele can be described as occurring in the population as a proportion, p, of all varieties of a specific gene. The change in p, where we indicate change using the Greek letter capital delta, Δ, is Δp (read out loud as “delta-p” or “change in allele frequency”). This gives a quick shorthand for measuring evolution: Δp 0. If the proportion of an allele changes in a population over generational time, then we have evolution in action. Game on!

To be fair, here, we ought to use the same sexy mathematical notation for phenotypic change. The population’s average (what statisticians call “mean”) value of a trait is represented mathematically by (read out loud as “X bar” or “mean of the trait”). As long as this trait, like the length of a finch’s bill, has some genes that determine it, then we have another shorthand for measuring evolution: Δ 0. If the mean of a phenotypic trait changes in a population over generational time, then we may also have evolution in action.

To prepare you for the robotic world you’ll encounter in the upcoming chapters, I should mention at this point that one of the great things about creating your own evolutionary world is that you get to do things like predetermine how genes relate to phenotype. Rather than having to worry about how heritable a phenotype trait was, we just decided that genetics would control entirely every trait, X, and every variation of X. Thus, any phenotypic changes that we might see in a population would have a direct and proportional genetic underpinning: Δ = Δp.

Isn’t that tidy? To be careful about what we have wrought, we would say that in our population of robots any phenotypic change equals a proportional genetic change. Tidy, indeed.

There is a problem with all of this, a perceptual one: usually the Δ from generation to generation is so small that we, as observers, don’t recognize the changes. The goldfinches in my garden this year look just like the goldfinches in my garden last year. We are blind to slow and steady changes, even those that happen over the course of a few minutes right in front of us. This phenomenon has been called “change blindness,” and the fact that it happens predictably is a startling testament to the fact that we have to be told to pay attention to most things in order to notice when they change. Misdirect attention and you have a magic trick. Thus, it’s no wonder that we don’t automatically track the evolutionary changes happening around us all the time. For example, unless you’re a gardener, you are unlikely to have noticed the Oriental bittersweet vine that has slowly crept into the shrubs and bushes of your midwestern and northeastern US yards since it was introduced in the 1860s.

Fortunately for us, Darwin—having trained with the best naturalists of the day, having traveled the world collecting samples, and having bred pigeons—was well placed to see variation and change on a small scale. Combined with his knowledge of Charles Lyell’s geology, he knew that the world was old enough to have let that kind of variation build up over time to become the huge changes that differentiate whales from hippopotamuses or tuna from trout. In today’s parlance microevolutionary changes cause macroevolutionary changes.[6]

Most biologists looking to measure evolution tend to focus on specific traits, or characters. John Lundberg, a biologist famous for his work on the evolutionary relationships of catfishes and the freshwater fishes of South America, told the graduate-student version of me that a character was any feature of an organism that you can observe or measure. In practice, then, you end up counting the number of spines in the dorsal fin in a population of bluegill sunfish and pumpkinseed sunfish. Or you measure whether or not the males in each species make and defend nests on the edges of the lake. Or you sample and sequence DNA to compare the alleles that make the little colored flap that sticks off the back of the gill cover. The result is that we tend to focus our analytic efforts not on the evolution of the population or species but on the evolution of one or two traits.

We do this even though we know that selection does not compartmentalize traits: the “whole animal interacting with the world and creating behavior” is really what is being selected at any given time and place, so some traits evolve not because they are the specific target of selection but because they just happen to be part of the whole animal. Changes in some traits may help the animal play the game of life whereas changes in other traits may hinder. Some changes may be neutral, but if selection on one trait is strong enough, the rest just get dragged along for the ride. For now, however, we’ll just think about traits as isolated evolutionary units. To do this oversimplification, we have to perform the convenient assumption called ceteris paribus, Latin for “all else being equal.” Under ceteris paribus thinking, we pretend that when we change the one thing that we are interested in, like a trait, nothing else changes or is influenced by that change. The logic of ceteris paribus is that we isolate one variable and understand how it influences the behavior of the whole system.[7]

We use ceteris paribus thinking all the time: eliminating one kind of food at a time to see if we have allergies, trying high-octane gasoline in our car to see if that improves mileage, altering our posture to see if that makes our back feel better, or testing a new drug for the treatment of multiple sclerosis in a clinical trial. Ceteris paribus is a great approach if all other variables remain constant and you have the discipline not to change other variables at the same time. If you remove wheat and dairy from your diet together, and that muscle soreness disappears, then you still have to go back and test each separately to know which one is causing the problem—or if it is the interaction of the two.

Using ceteris paribus, then, we can ask if any single trait is an adaptation. This is the equivalent of asking if a trait has evolved because it was the target of natural selection. Keep in mind that this use of “adaptation” as a noun is different from the verb of “adapting,” which refers to the process of natural selection in action. In addition, if we are being careful, we’ll always ask if a trait is an adaptation for a specific situation.

To answer this kind of question, we need information, and lots of it. Fortunately, Robert Brandon has carefully analyzed the kinds of information that are necessary and sufficient to provide what he calls a “how-probably” explanation of adaptation (Figure 2.2).[8] A how-probably explanation of adaptation, by the way, is rarely accomplished because we usually are missing one or more pieces of evidence. We miss loads of evidence when we are dealing with adaptation in extinct life-forms, and then the best we can do with our partial set of information is to claim that we have a “how-possibly” explanation.

The evidence that we need to test a how-probably hypothesis of adaptation begins with understanding the trait of interest. First, we need to know that the trait is heritable, how it is genetically coded, and how it interacts, at the level of DNA, with other heritable traits. You can see how this evidence fits in with the definition of natural selection from earlier in this chapter. Second, we need to understand “polarity” of the trait—that is, what did the trait evolve from? What did the trait look like in its ancestral form, and what does it look like in its derived form? In the case of a jointed vertebral column, we know that it evolved from an unjointed notochord. Third, we need to understand how the ancestral and derived forms of the trait—and all the intermediate forms in-between—functioned in a living individual.

We also need information about the population in which the trait is evolving. First, we need to know about the structure of the population, things like number of individuals, age at sexual maturity, and rates of immigration and emigration, to name a few. Second, we need to know about what Brandon calls the “selection environment”—what I think of as the world in which the population exists. This world includes both physical and biological factors. Most importantly, the world contains other individuals very much like you, and because of that similarity, you are likely to interact and compete with those other members of your population. All of these features in the world make up the “selection pressure.” Third, we need to know how the population responds to selection. This gets us back to how we measure evolutionary change, with Δ and Δp.

If you can muster all of that information, you have what Brandon considers to be an “ideally complete” explanation of adaptation. But you can see the problem with these how-probably explanations: you basically need to know everything there is to know about the trait and the population! This is what makes the Grants’ work on the ground finches in the Galapagos so impressive: they have over twenty years of data on the genetics and function of multiple phenotypic traits and over twenty years of data on the demography, selection environment, and responses to selection of the population of medium ground finches on the island of Daphne Major.

Keeping Brandon’s necessary and sufficient information in mind (Figure 2.2), you can see that one of the brilliant decisions that the Grants made was to select a population that was isolated (very little immigration and emigration), small, and in a simple selection environment (open habitat with only a few other animal and plant species). As Wake Forest University’s David Anderson, another bird expert working in the Galapagos says, the birds on those geologically new and ecologically simple islands suffer out in the open.

What Anderson means by “suffering out in the open” is that humans who spend the time to observe carefully in the Galapagos can actually watch many events that have huge evolutionary impacts. For example, Anderson watches in lean years as Nazca boobies can only produce a few or feed some of their chicks. Reproductive success or failure is there, out in the open, for him to observe.

Make babies and help them make babies. If you are a Galapagos finch and you do this better than other Galapagos finches, then you are a winner in the game of life. Your score is based on how well you do relative to others in your population. If you are the best, you get a score of 1.0. If you are the worst and don’t produce any offspring, you get a score of 0.0. This score is called your “evolutionary fitness.”

Scoring the game of life is just the beginning. Once you have the score, the natural question to ask is, why do some individuals play the game better than others? And then, what about the individual and its interactions with its world matter? When you can answer these questions, then you’ve got a handle on which traits are important, how those traits function, what in the world selects individuals, and how the population responds to those selection pressures.

Anderson and the Grants were both lucky and smart—they managed to find an environment in which this scoring is, if not easy to do, at least possible. Most biologists don’t have this advantage. Thanks to our decision to study evolving robots, my colleagues and I suddenly found ourselves in a position a lot like that of the biologists studying Galapagos finches: we could watch a population that suffered out in the open. We can create our own simplified world, create individuals whose genetics we know, create a population whose structure is predetermined, and then carefully observe behavior and evolution as the individual robots interact with their world. Because we also set up what is called the “fitness function,” we are also the judges of the behavior of individuals. We become the agents of selection.

The idea of evolving robots is not new to my laboratory. Stefano Nolfi and Dario Floreano brought the concept to the general academic world with their book, Evolutionary Robotics, which was published in 2000. From the context of artificial intelligence, cognitive science, and engineering, they helped create a framework in which researchers could harness evolutionary processes—randomness, selection, and differential reproduction—to create without their guidance new kinds of behaviors and intelligence in mobile robots.

What we’ve done is to take Nolfi and Floreano’s evolutionary robotics framework and apply it to biology (Figure 2.3). Whereas Nolfi and Floreano weren’t originally trying to build biologically realistic robots, that’s where we start. And the inspiration for that approach came from Barbara Webb, an invertebrate neuroscientist and behaviorist who figured out that she could use robots to test hypotheses about the neural underpinnings of animal behavior.[9] When this approach—using physical robots to test hypotheses about biological systems—is thought of in general terms, Webb calls the field biorobotics. The combination of these two approaches creates evolutionary biorobotics.

So if we’re going to build robots that can really play the game of life, they must be able to reproduce, have behaviors and other traits that are genetically heritable, and have limits placed on the number of offspring that can be reproduced. Putting these features into a robotic system gives us what we like to call the lifecycle of evolving robots (Figure 2.4).

To be frank, evolutionary biorobotics has four important limitations when it deals with extinct species and their evolution. First, as we discussed earlier when talking about the kinds of evidence that you need to explain an adaptation (Figure 2.2), analyses of past selection are fraught with potentially crippling and untestable assumptions about the genetic structure of the population; the genetics of traits in question; and the pattern, strength, and phenotypic targets of selection. Second, what you can reconstruct and test is only the ecological function of the character, the selection environment, and the response of the population to selection. Third, because we create model simulations with our robots, our reasoning is by analogy. So as we set out to explore the evolution of backbones in robotic fish, the best we could hope for was robust support—in digital and embodied populations—for the prediction that selection for swimming abilities drove the evolution of the backbone in real fish. In the worst case, the best we’d be able to say is the obvious: that different selection environments can produce different results in different robot-world systems. Fourth and finally, our use of digital and embodied robots interacting in constructed worlds grossly simplifies the animal, its environment, and the animal-environment interaction.

Still, there is much to be excited about: at the minimum, if varying our robotic backbones changes robotic behavior, at least we’d have a proof of concept that we were studying an important variable that may or may not have been under selection at some point. Second, the fact that robots evolve can give us insight into how the process of adaptation works, whether in robots or biological organisms. And at least we knew we were in good company: model simulations with digital agents have already been used, most notably by Charles Ofria and Richard Lenski at the Digital Evolution Laboratory at Michigan State, to test a range of biological hypotheses about evolution.

That left us with the task of designing our first biorobots. Let’s engineer some players for the game of life.