Darwin’s Devices

Chapter 4 TADROS PLAY THE GAME OF LIFE

TADROS EVOLVE! BY ANY MEASURE, OUR POPULATION of autonomous, aquatic Evolvabots successfully played the game of life (Figure 4.1). No surprise, I hope—indeed, the surprise would have been if they didn’t evolve, as that was what we designed them to do.

If you begin at the beginning and then go on until you get to the end,[31] you would see that from the design of Tadro1 in 2003 to the publication of our Tadro3 paper in 2006, our Tadro team had more than twenty-three members. Okay, so that’s not a small-to mediumsized village. But it’s still some measure of how hard this work is.

Kira Irving, Keon Combie, Virginia Engel, and Joe Schumacher led a group of Tadro3 operators and biomimetic tail benders that included Nicole Doorly, Yusuke Kumai, Gianna McArthur, and Kurt Bantilan. They performed 120 trials over ten generations, building 360 tails with biomimetic backbones along the way. Each trial also had three minutes of videotape to analyze in one-second increments; in each frame, we had to mark (and then double-check) the position of the light and the bow and the stern of each Tadro3. We used these points to compute average speed, average wobble, time to the food, and average distance to the light source. Twelve of these numbers for each individual tail phenotype (each phenotype was described by two traits: length of tail, L, and the biomimetic notochord’s material stiffness, E—more on these later) were then used to calculate the three individual fitnesses corresponding to each phenotype—fitnesses that we needed to run the genetic algorithm that spat out the next generation’s phenotypes. When we were really cranking full time in the summer—including evenings—a group of four or five of us could get a generation done in about four days. But if anything went wrong, like when we dropped a Tadro upside-down in the water, then we’d have to stop and rebuild before we could get back to the evolutionary trials. All told, the trials took ten weeks.

As we let our population of Tadro3s play its game of life, we found that the feeding behavior, tail stiffness, and genetic composition of the population all changed over generational time (Figure 4.1). Although the fact that the population evolved wasn’t a surprise, we were surprised about the direction evolution took. I should say directions, plural. Under a steady selection pressure that rewarded enhanced feeding behavior, we expected steady, directional change. Instead, what we see, even in this simplified Tadro3 world, is that evolutionary change oscillates, moving in different directions at different times.

Given that we know the Tadro3 world inside and out, we sure as heck-fire better be able to explain the unexpected directions in the evolution of the population Tadro3s.

Let’s jump-start the process of interpretation by revisiting the big picture. We were interested in creating Evolvabots that could test a hypothesis about the evolution of early vertebrates: natural selection for enhanced feeding behavior drove the evolution of vertebrae in early vertebrates. From this hypothesis we came up with a primary prediction: selection for enhanced feeding behavior will cause the population of Tadro3s to evolve stiffer tails. Implied in that prediction is another: the evolutionary change will be directional, moving the population from flexible to stiff tails in concert with ever-improving feeding behavior.

What’s clear is that these predictions are too simple. Take a close look at the reality of the hard data (Figure 4.1). The average score for feeding behavior[32] increases greatly from generation 1 to 2 while the standard deviation,[33] represented by the length of the bars emanating from the filled circles, decreases. This initial change sure looks like the direction we predicted: enhanced feeding behavior. The decrease in variance might be expected, too, because selection was just picking individuals with the highest feeding behavior to reproduce. But in the very next generation, 3, the average score for feeding behavior drops and then continues downward through generation 4. What’s up, Doc?

At first glance, this downward trend looks crazy-wrong: how can we select for improved feeding behavior and get just the opposite? The short answer is this: in generation 2 differences in feeding behavior among individuals were not large enough to cause differences in reproduction. All three individuals contributed the same number of gametes (egg and sperm)—two—to the mating pool of six gametes. Our mating algorithm assigned differences in reproductive output according to the differences in fitness between individuals. Even though the fitness function assigned slightly different numbers to each individual—based on their abilities to swim quickly, reach the light target quickly, stay and feed near the light target, and move around smoothly—in generation 2 those fitness differences were just too small to matter.[34]

When individuals contribute equally to the next generation, we have an evolutionary tie. This tie means that the parents are likely to make a generation of offspring that looks, on average, like themselves. In evolutionary terms, there was an absence of selection, or no selection pressure. Either phrase might sound a bit inaccurate, as we had selection judging individual Tadro3s in a given generation using our fitness-function scorecard. But keep in mind that ultimately it’s differential reproduction among individuals that makes evolution by natural selection work.

In the absence of selection, how, then, do we get evolutionary change? Recall Darwin’s lament from Chapter 2: “Great is the power of steady misinterpretation.” He was referring to the fact that many scientists overplayed the power of selection, to the extent that they ignored other evolutionary mechanisms and, as a result, saw adaptations in every cranial bulge and fingerprint, when sometimes randomness is what’s at work.

Darwin had little hard evidence to counter this misconception because he lacked our understanding of genetics, which, ironically, was developed by a contemporary of Darwin’s, Gregor Mendel, but languished in an obscure journal until the early twentieth century. With our current understanding of genetics, of course, we know that random genetic changes are always occurring, both in the germ-line and other cells. More randomness can come into play during mating. Although some mating is decidedly not random, for many organisms it is. We set up our Tadro3 population to engage in random mating. After each gamete was given a certain probability of mutating or not, we took all six of the gametes the parents produced, and combined the gametes into pairs randomly. Those pairs of haploid gametes combined to give us the new diploid genome for each new baby Tadro3.

At this point, my colleague Rob Root, a mathematician and central collaborator on many of our Tadro projects, would want me to remind you that we have a problem with our randomness because we are operating in the realm of “the mathematics of small numbers.” Our population size of three is simply too small for the statistical assumptions of flip-a-coin randomness to hold. The result, if you were flipping coins, is that you could easily hit heads, heads, and heads, three in a row. You’d say that you were “on a roll,” and it sure would not look like a random process until you flipped the coin about twenty more times. Geneticists call this mathematical description of what happens in populations of small size “genetic drift.”[35]

Genetic drift, in the absence of selection, produces evolutionary change in directions that are random with respect to both phenotype and genotype. Because drift is random, neither it nor any other random mechanism will produce the kind of long-term directional pattern that we recognize as the signal of adaptation. Only selection is equal to the task.

For our tiny population of Tadro3s, the absence of selection in some generations allows the combined random chance of two sources of randomness—mutation and genetic drift—to be the prime movers. Over a generation or two, we can even be fooled. Our population can be “on a roll” that may look like a directional and evolutionary trend. This is what happened to the average feeding behavior from generations 2 through 4. Random changes just happen to have combined to decrease the feeding behavior. Chance favors no one.

What’s really important to keep in mind, and to keep Darwin from being on a roll in his grave, is that over generational time random changes—which occur with or without selection—are, indeed, mechanisms of evolutionary change. This fact is worth restating: you can have evolution with or without selection. We’ve shown this to be true with our population of Tadro3s. It’s also worth pointing out that the random genetic processes are always occurring and that they operate independently from selection. There is a third mechanism influencing how populations evolve, however: history (Figure 4.2).[36] The independence of these three kinds of mechanisms was beautifully demonstrated by Professor Rich Lenski, of Michigan State University, in his work first on bacteria and then on what he calls “digital organisms.” Lenski showed that the genetic and phenotypic variation that exist in your population at any time constrain the population’s future evolutionary possibilities. Any finite population can only evolve in certain directions, directions that are constrained by the underlying genetic coding of the phenotypes and the responses of individual organisms to the particular environment. Another way to think of this history effect is that selection can only select, as we saw in the Tadro3 generations 2 through 4, when individuals vary. Selection can only result in evolution-by-selection when genes code, at least in part, those individual variations.

For our population of Tadro3s, history covers a multitude of sins or, more accurately, assumptions. Our Tadro3s didn’t first evolve from single-celled Tadro0.001s, so they don’t have an explicit evolutionary history, but they do have an implicit one: the evolutionary history of the tunicate tadpoles after which Tadro3s are modeled. That tunicate history has bequeathed Tadro3s a notochord, a light-sensitive organ, a brain capable of linking sensor and undulating tail, and the genes that underwrite these phenotypes. History has also brought along environmental baggage—the water with a directional light source in which Tadro3 lives.

History’s most important legacy for Tadro3s is the blood line, the initial genetic conditions—essentially the genes and the variants of those genes that we decided Tadros had. Because we were interested in structural stiffness of the notochord and, by extension, the tail in which the notochord sits, we coded for that property and placed its initial value in the middle of a scale of biological stiffnesses.

Structural stiffness is a property that describes how a structure will change shape when an external force acts on it. Think about a cantilevered structure, like a flagpole hung horizontally. When you put a flag on the pole, it bends down slightly. Or, better yet, imagine that old Three Stooges movie short, Flagpole Jitters, in which Moe, Larry, and Shemp find themselves hypnotized and walking out on a flagpole high above the city streets. Through experiments with weights rather than clinging Stooges, engineers have figured out that the structural stiffness of a cantilevered beam like a horizontal flagpole is proportional to the beam’s flexural stiffness and inversely proportional to the cube of the beam’s length. This is one of those moments when, against the advice of my friends, I just can’t help but use an equation to summarize all of this:

To clarify, what this equation says is that structural stiffness, represented by the variable k (in units of Newtons per meter) is defined as the ratio of the flexural stiffness, the composite variable EI (in units of Newton square-meters), to the cube of length, L³ (units of cubic meters). What’s nifty about this equation is that you can see right away what matters. Want a stiffer beam? Increase the EI or decrease the L. A longer cantilever has a huge impact on k because of the cubing of the L. An equation like this also helps conjure up the genetics we’re after.

We could have just said that Tadros have genes that code for k directly and then left it at that. However, research on biological stiffness suggests that all three variables, E, I, and L can change independently during development and evolution. The variable E (in measurement units of Newtons per square meter) by itself is called by a variety of names: “modulus,” “elastic modulus,” “complex modulus,” “Young’s modulus,” or “Young’s modulus of elasticity.” Too many aliases! But wait—let me do my part for the witness protection of E. Because E is part of structural stiffness, k, and flexural stiffness, EI, and is caused by the kind and number of chemical bonds in the material you’re dealing with, I prefer to call it “material stiffness.”[37]

To allow for the likely possibility that selection might target the length of the tail for a variety of reasons, some of which may have to do with structural stiffness and some not, we created a genome that coded for L and E separately. By not coding for the variable I, we were holding that part of the geometry—and everything else about Tadro3 for that matter—constant. In the language of a geneticist, both L and E were quantitative characters, polygenes, multiple loci capable of producing smooth gradations in the phenotypes for which they code. All loci for L were located on a chromosome separate from the loci for E in order to allow for independent assortment. In other words, having quantitative traits means that the genome does not contain the simple on-off, wrinkled pea or smooth, kind of genes that we call “Mendelian.” Each set of genes is, instead, a continuous number scale, capable, within a given window, of producing a range of E values different from a range of L values.

You can see the independent changes in the proportion of E and L genes in the bottom panel of Figure 4.1. Notice that as the proportion of L genes increases from generation 7 onward, the structural stiffness, k, in the middle panel, plunges. This is exactly what we’d expect from our equation for k, on previous page. That L³ term in the denominator is increasing dramatically, and it is lowering k at the same time that the E term in the numerator is decreasing and also lowering k. Faced with this kind of genetic evolution, poor old structural stiffness doesn’t stand a chance.

Over the course of ten generations, the population’s average value of the structural stiffness of the tail, k, plummets from above 5 to below 1 Nm^-1. We’ve seen what was happening genetically to cause the decreased value of structural stiffness. But these genetic changes don’t speak to how selection—which judges individuals by their behavior, not by their genetics—was interacting with randomness and history. We still have two bothersome itches to scratch: (1) Why did the structural stiffness decrease under selection for enhanced feeding behavior when we predicted that it would increase? (2) Why does feeding behavior seem at times to be unrelated to the structural stiffness of the notochord?

I want to warn you right now about a tempting siren who begins singing on the rocks at about this point in a study. When, as was the case with Tadro3, your experiment produces a result that appears to be the exact opposite of what you predicted, the immediate emotional response is to be disappointed and self-flagellating. My students and I certainly were. When we graphed the data in Figure 4.1, we had to have a group counseling session immediately to air concerns and responses. In the lightning reaction round, we heard: What went wrong? Our experiment didn’t work! These data suck! We stink as scientists! My line then, and I’m sticking to it now, is that if you design an experiment carefully, execute it well by tracking down mistakes as they occur and running controls, your data will always be great. Data just are. No matter what those data say about your predictions, they and the experiment that generated them stand on their own, with their total value determined by how well you measured what you set out to measure.

The negative emotional response, I reckon, comes from the fact that we all secretly think that we understand our experimental system well enough to know exactly how it will turn out. We are, from an emotional point of view, just running through the experiment to show other people what we’ve already figured out in our heads.[38] By the time we’ve made our prediction, a process that is really like running our own internal model cognitively, we have committed emotionally to a particular outcome. We aim to “prove” our point through demonstration.

Although disappointment and disillusionment in the face of unexpected results may be natural emotional responses—and ones that I share with my students—they run counter to the way that many but not all of us reason scientifically. Strictly speaking, we demonstrate that some testable concept is true through our repeated failure to show it to be false.[39] Although we can certainly demonstrate that something predictable happens every time we release our coffee mug from a height of two meters, no one has seen gravity.[40] Gravity is a concept for a kind of energy related to the masses of objects. The relationships that we see between objects on planets and between planets and stars in space is observable and consistent with our concept of gravity; hence, having failed repeatedly to disprove those consistent relationships between objects, most of us think that gravity is a fact. Because the failure to reject gives us confidence in the truth that we infer about gravity, we employ gravity, in turn, as an inferential tool to create new scenarios. We use the fact of gravity to infer the presence of the universe’s unseen dark matter. If dark matter were shown to be nonexistent, that observation would indirectly refute the gravitational paradigm. In sum, to prove, we attempt to refute.[41]

Following our therapy session, which included a harnessed descent into the cave of refutational negativism, our Tadro team decided to look deeper into the data. We needed to figure out if, in fact, our data sucked (which is always a real possibility) or if the results were screaming in our faces about something really interesting that we just hadn’t anticipated. I’ll spare you the tedium involved in figuring out if your data suck: it’s all the usual kinds of things about checking your transcriptions, lab notebooks, mathematical formulae, control experiments, and calibrations of instruments.[42] We came to the conclusion that we couldn’t explain our results away with the reflexive “bad data by bad scientists.” Something much more interesting was afoot.

If you look back at Figure 4.1, you’ll notice that from generations 1 to 2, when selection was present, we had a big increase in the feeding behavior score that was correlated with a jump in structural stiffness. This pattern was as we predicted, and it allowed us to make the nifty overlay in Figure 4.3 in which the Tadro3 with the stiffest tail has the best feeding behavior and the Tadro with the most flexible tail has the worst. Also from generations 2 to 3, when selection wasn’t present, we still see that feeding behavior and structural stiffness are correlated, changing together. All is well, even though stiffness is dropping a bit. No big surprise, given that random genetic changes can be dominant in the absence of selection, as we talked about earlier.

After generation 2 the system appears to run amok: from generations 3 to 4 and beyond, really, the one-to-one connection between behavior and stiffness disappears. Behavior drops or stays constant while structural stiffness increases. Or the behavior score increases while structural stiffness drops, as in generations 4 to 5. In this case, in the absence of selection we know that we have only random genetic processes driving the evolutionary changes. Randomness challenges our assumptions. Because of the one-to-one relationship between behavior and stiffness from generations 1 to 3, we assumed that the structural stiffness of the axial skeleton was causally related to the feeding behavior. But faced with the evidence from later evolutionary changes, that relationship is either not true or it’s only true some of the time. How can we tell?

When complexity dims the light of interpretation, one way to navigate is to stop and examine your assumptions. In our case, the first assumption we tested was a fundamental one: when we selected for enhanced feeding behavior the population responded by evolving enhanced feeding behavior. This was what happened. Thank goodness! Selection was present in four generations: 1, 5, 6, and 9; in three of those cases (generations 1, 6, and 9), the ensuing generations showed higher average feeding scores than their parental generation (see also Figure 4.4). We took the data from each individual, not just the averages, from all four of those generations with selection and statistically tested the mean response to selection. The statistical test confirmed what we see by eye: on average, selection evolves enhanced feeding behavior. Keep in mind that even when selection is acting, randomness will almost always deflect, to lesser or greater degree, the population’s evolutionary trajectory that selection proposes (see Figure 4.2).

We can learn something new (for us) and important by examining the only time, from generation 5 to 6, when selection didn’t evolve improved feeding behavior. If you look at the change in genes that results from selection (bottom panel, Figure 4.1), you can see a drop in tail length, L, accompanied by a jump in material stiffness, E. Going back to our handy-dandy equation (I just knew it would be useful) for structural stiffness, k, we know that because of the magnifying effects of the L³ term in the denominator, the population’s average k must be higher in generation 6 than in generation 5. And, indeed, we see that jump in average k in the structural stiffness plot just above the genes plot. This connection of genes and increased structural stiffness rules out the random genetic deflection idea in this case as the primary cause of the evolution of the feeding behavior. To explain this fully, though, we have to go on a little journey. Fasten your safety harness, please.

What the disconnect between selection in generation 5 and feeding behavior in generation 6 suggests is that we need to test this assumption: feeding behavior is causally connected to structural stiffness of the notochord. If this assumption were always true, then we’d expect to see behavior improve or decline in concert with increases or decreases, respectively, in the notochord’s structural stiffness. From our previous discussion, we know that behavior and stiffness don’t show any regular pattern that might lead us to believe that they were causally connected. However, several other patterns are possible. First, it could be that changes in the notochords’ structural stiffness are correlated not with the overall feeding behavior but rather with some of feeding’s sub-behaviors: swimming speed, body wobble, average distance from the food, and time to find the food. Second, those sub-behaviors might not be correlated with structural stiffness but rather with the stiffnesses’ subcharacters: material stiffness, E, and length of the tail, L.

We ran a series of statistical tests to look at the patterns of correlation of the stiffness variables on one hand—structural stiffness, k, material stiffness, E, and length of the tail, L—and the behavioral variables on the other—swimming speed, V, body wobble, W, distance from food, D, and time to find the food, T. As separate independent variables in a linear regression, k, E, or L all predict about 20 percent of the variation in V and W but predict none of the variation in D and T.[43] Moreover, k and E are positively correlated with V and W, and L is negatively correlated with V and W. Thus, we’ve got a clear relationship between structural stiffness of the notochord and swimming speed and body wobble, two of the four components of the feeding behavior score.

What’s strange about this pattern of correlations is that swimming speed and body wobble are positively correlated. Recall that we had predetermined that the fitness function would judge increased speed as “good” and increased body wobble, distance to the food, and time to the food as “bad.” Thus, the fitness function should, when selection is strong enough to create differential reproduction, create this specific pattern of correlations among the sub-behaviors (Figure 4.4). Instead, the evolutionary pattern that we always got was the one we just identified with statistics on individuals across generations: speed and wobble always increased together under selection.

Surprise! When we select for improved feeding behavior we actually make the evolutionary fitness of the Tadro3s better and worse at the same time. Tadros swim faster (and that’s good for an individual’s fitness score) but they wobble more (and that’s bad for the fitness score). To explore what was going on with this odd couple, we swam our Tadros with all their different tails in a simple swimming trial—no competition, no evolution. Just swim straight down the tank. We videotaped the Tadro3s and measured their speed and body wobble. Under these conditions speed and wobble were not correlated. In other words, in the “wild,” competing with other Tadro3s to swim and feed, speed and wobble were, for some reason, functionally connected. But in the “lab,” swimming by themselves in a straight line, Tadro3s didn’t show this functional connection.

All is revealed when we look closely at this sub-behavior that I’ve been calling body wobble. You may recall from the previous chapter that we said that wobble was a measure of how unsmooth the swimming path of the Tadro was. If you ever read Arthur Ransome’s Swallows and Amazons series as a kid, you’ll know that you can spot a novice helmsman on a small sailboat by the wiggles in the boat’s wake. Those wiggles, which come from an unsteady hand on the tiller, waste energy by rotating the boat back and forth, slowing down the boat’s forward progression. We took Ransome’s wiggles and turned them into wobbles. From videotape records, we measured the change in heading of each Tadro3 every second during the three-minute trial. We then calculated, over every two seconds, how rapidly the change in heading was occurring. If you are a sailor who navigated high school physics successfully, you may recognize this as a measure of angular acceleration in yaw. We then took all of the angular accelerations for the whole trial and calculated their standard deviation, which is a measure of how variable the angular acceleration was. This gave us a single number called wobble (in units of radians second^-2).

Check it out: wobble not only measures the rate at which the Tadro is wiggling and losing energy, what we call recoil from the flapping tail, but it also measures the presence of high-speed maneuvers. Think of it this way: a quick turn is a big, quick wiggle. We figured this out by running some digital simulations in which we would make a swimming Tadro perform a big turn quickly. We measured the resulting increase in wobble. It turns out that frequent turning maneuvers produce a huge amount of wobble, much more than the energy-wasting recoil.

We realized that wobble in the wild was picking up on all of Tadro3s quick turning maneuvers that weren’t present in the lab’s straight swimming. Cool! We had accidentally measured something different and even more interesting than what we’d intended. Now it made sense why swimming speed and wobble were positively correlated in the wild: if you are swimming faster you can turn faster. Plain and simple. Every driver and sailor knows this and slows down coming into a turn in order to maneuver smoothly or, instead, maintains their headway to make a break-neck turn.

In Tadro3s we now see that body wobble is functionally dependent on swimming speed. This gets us into a really interesting area in terms of evolutionary biology. Any trait that has a genetic basis and whose phenotypic expression is influenced by a different gene is said to be “epistatic.”[44] Epistasis is a very important genetic phenomenon and can occur in a variety of ways that biologists are still uncovering. With Tadro3s, though, the interaction is not conducted directly at the genetic level; we even set up the genes to make sure that they didn’t interact at the level of the genome. Instead, here we see phenotypic epistasis, a physical interaction of the sub-behavior phenotypes that occurs because the phenotypes share a single body. We established the genetic basis of speed and wobble through their connection to structural stiffness of the notochord, which is directly coded by the Tadro3s’ quantitative genes.

Now that our wobbly journey through epistatic seas is over, we can stand on terra firma when we explain the evolutionary change, under selection, from generation 5 to 6 (see Figure 4.1, again!). We had observed that this was the only time, out of four evolution-by-selection events, when selection made swimming behavior worse. We had established the fact, by looking at the changes in the proportion of genes, that random genetic effects couldn’t explain the evolutionary change. What was left to analyze was the functional relationship between the notochord’s structural stiffness, which is coded genetically, and the feeding behavior, which is judged by selection. One of the sub-behaviors, body wobble, turned out to be beneficial to improved feeding behavior, instead of detrimental, as originally thought. In addition, body wobble is functionally linked to swimming behavior.

These two final and unexpected facts about body wobble explain the also-unexpected degradation of feeding behavior from generations 5 to 6. Because the fitness function rewards increased swimming speed while penalizing increased wobble, the composite score of feeding behavior drops. Keep in mind that we measured feeding behavior using the same relationships that we used in the fitness function.

Knowing what we know now, this was a mistake. We should’ve rewarded increased wobble and called it something more appropriate, like “agility.” Unfortunately, we can’t go back and change evolutionary history without completely redoing the experiments (more on why in the following pages). We calculated the fitness function every generation in order to make the next generation. But we can recalculate the feeding behavior score with increased wobble rewarded, not penalized.

When increased wobble is rewarded, we get a slightly and importantly different picture of the evolution of the composite behavior that we call feeding of Tadro3s (Figure 4.5). First, the averages of the population’s feeding behavior now undergo more change, moving both higher and lower than their means under the old fitness-based measurement scheme. Second, one of the generation-to-generation evolutionary changes is different: from generations 5 to 6, under selection, average feeding behavior now improves rather than degrades. In other words, armed with our new understanding of wobble as a positive metric of rapid maneuvers, feeding behavior actually did improve under selection!

This seems to be a most ingenious paradox. We are saying that feeding behavior declined from generations 5 to 6, but really, it improved. What? Allow me to sum up. Before the evolution of Tadro3s started, we created a fitness function that we thought was selecting for improved feeding behavior. This fitness function rewarded increases in swimming speed and penalized increases in body wobble, time to find the food, and distance from the food. In each generation, if individual Tadro3s varied enough in their feeding for the mating algorithm to create differential reproduction, then selection was both present and active as an evolutionary mechanism.

In only one case, generations 5 to 6, was selection present when average feeding behavior declined. We had measured the feeding behavior of individuals using the same four sub-behaviors and their goodness or badness as used by the fitness function. All that we changed in making the behavioral metric was to compare individuals not just in a generation but instead across all generations as well as to scale individual differences relative to the variance exhibited by all Tadro3s across all generations. Under this original fitness-based metric, we found that one apparent anomaly in feeding behavior.

We ruled out random genetic change as the cause of the decrease in feeding behavior from generations 5 to 6 because the proportion of genes changed to increase structural stiffness of the tail, and structural stiffness of the tail is positively linked to swimming speed. That left us to reconsider the four sub-behaviors and their functional interactions. We discovered that in the competitive arena, increased wobble wasn’t a sign of energy inefficiency but rather agile turning maneuvers when facilitated by high swimming speed. Hence, our original idea about what constitutes good and bad feeding behavior was just plain wrong.

When we altered our after-the-fact behavioral metric, we found that feeding behavior always, in fact, improved under selection. The paradox arose because the selection was actually penalizing increased wobble when, as we know now, it should have been rewarding it. As I mentioned earlier, we would love to revise the fitness function and repeat the experiment. We’d expect that selection, when present, would be even stronger, and that the jumps might be greater. But we have one big problem with repeating the work: evolution of physically embodied Evolvabots takes loads of time and buckets of money. This is one reason you try to be as careful as possible in the design phase (see Chapter 3)!

We thought that our Evolvabot design, carefully laid out as a series of simplified representations of nature in the previous chapter, would produce a simple evolutionary pattern. We could not have been wronger.[45] We evolved two phenotypes, material stiffness, E, and length of the tail, L, that together are responsible for the structural stiffness, k, of the notochord. These traits were coded as quantitative genes housed in a diploid genome. The possessors of the tails whose phenotype was dictated by the genes competed for food. Selection, codified by our fitness function, rewarded individuals in a particular generation who behaved better in terms of increased swimming speed, decreased body wobble, decreased average distance to the food, and decreased time to find the food. For reproduction, haploid gametes were mutated and combined in a simple random mating scheme to produce the instructions for the notochords of the next generation.

Our first surprise came when we saw, after ten generations of a constant selection pressure, that the evolutionary changes in the population’s feeding behavior, tail stiffness, and gene proportions were anything but constant. Why would a constant selection pressure produce different results each generation? Part of the answer is that in each generation the other agents in the world have changed, and their different evolved behaviors alter the competitive landscape. Another part of the answer is that genetic variability contracts and expands over generational time, changing the phenotypic options available for selection to judge.

Our next surprise came when we realized that selection was only operating to produce differential reproduction in four of the ten generations. This meant that in the generations without differential reproduction the evolutionary changes in phenotype and genotype were happening because of purely random effects. In particular, mutation and genetic drift caused mutational differences and individual genomic differences to combine into relatively large effects when selection was not present.

Our final surprise came as we probed the causal connections between the structural stiffness of the notochord and feeding behavior. Feeding behavior was measured by the same sub-behaviors that we put into our fitness function—swimming speed, body wobble, average distance to the food, and time to find the food. When we correlated these sub-behaviors with structural stiffness of the notochord, we found that swimming speed and body wobble were positively and significantly correlated with structural stiffness, k, material stiffness, E, and length of the tail, L. Time and distance to the food were not. This meant that when time and distance were undergoing greater evolutionary changes than were speed and wobble, structural stiffness could be decoupled from feeding behavior. This situation was complicated by the fact that speed and wobble are positively correlated in terms of function but are negatively correlated in the fitness function. Hence, in terms of fitness, their effects would tend to cancel.

Now that we are confident in our understanding of the mechanisms and interconnections driving the evolution of Tadro3s, we can be confident in doing what we meant to do all along: not just create an evolutionary simulation but test a hypothesis about the biological system that the simulation represents. We proposed the hypothesis that natural selection for enhanced feeding behavior drove the evolution of vertebrae in early vertebrates. From this hypothesis we came up with a primary prediction that we tested: selection for enhanced feeding behavior will cause the population of Tadro3s to evolve stiffer tails.

Our data refute this prediction (see Figure 4.1). Hence, the hypothesis from which it came is also refuted. We found just the opposite of our expectation in some cases—selection for enhanced feeding behavior caused the population of Tadro3s to evolve more flexible tails (see Figure 4.1, generational change in stiffness following selection). If we buy our own argument that Tadro3s and their water world represent important aspects of the earliest vertebrates, we have to conclude that selection on feeding behavior was unlikely to have been the primary driver of the evolution of vertebrae. In science this kind of failure is called progress.

But if not feeding, then what drove the evolution of vertebrae? The positive relationship that we’ve shown between tail stiffness and swimming speed + turning maneuvers offer one alternative hypothesis: selection for speed and maneuverability alone drove the evolution of vertebrae. The problem is that in order for selection to work on locomotor abilities alone, it can’t be simultaneously working on other things like feeding behavior. If it does, we get these oscillating patterns of change that we saw in this experiment, when stiffness is sometimes correlated with feeding behavior and sometimes not.

This kind of complex response to selection, by the way, is very realistic in living fishes, as David Reznik of the University of California and his colleagues have shown.[46] Based on extensive review of the literature on responses of fish populations to selection, they contend that a behavior like acceleration performance is influenced by a network of traits, all of which, because of their genetic properties and functional connections to other behaviors, can be under countervailing selection pressures at the same time. Dominant selection pressures on a population can also vary in the wild, Reznik and his colleagues have shown, as predators move in and out of small pools containing mating subpopulations of Trinidadian guppies.

Predators appear to create strong selection pressures in other species of fish as well. Dohlf Schluter and his colleagues at the University of British Columbia have shown that moving threespine sticklebacks from ocean habitats to freshwater lakes—migrations that occur naturally—appears to release the immigrant population from predation pressures found in marine environments.[47] Without predation from other vertebrates, the sticklebacks respond genetically and, over generational time, grow faster and produce less body armor. Moreover, Richard Blob, of Clemson University, and his colleagues suggest that predation may have provided the selection pressure for stream gobies to evolve a remarkable ability to scale waterfalls in Hawaii.[48]

A different kind of interpretation of our results is that we didn’t really test a hypothesis about vertebrae because we had used the stiffness of the notochord as a proxy for the number of vertebrae. Even though this relationship is mechanically based, with stiffness and vertebrae being positively related, what if something else about vertebrae mattered? Perhaps the presence of vertebrae increases stiffness and changes the curvature of the tail, its shape, and the way it acts as a propeller?

Another valid criticism of our work with Tadro3 is that its brain was just too simple to adequately model even a simple system like the tunicate tadpole larva. But, we counter, the point is to create the simplest system possible because even the simplest autonomous agents produce complex behaviors and complex evolutionary patterns. You test your hypothesis using the simplified model. Then, based on the results of the test, you interpret what happened; you work to understand the mechanisms operating at multiple levels in your system. Interpretation, as we’ve seen, is darned tricky, even when you operate under the KISS principle.

What we’ve learned from Tadro3 is that neither the selection pressure, the design of Tadro3 itself, nor an interaction of the selection environment and the Tadro3 agent explain why vertebrae are likely to have evolved. We haven’t solved the puzzle!

So our next step is to think about how to make Tadro and the selection pressure a bit more complex. Thinking about our design principles from Chapter 3, we have to make sure that we understand the biology well enough so that we can understand what it is that we want in our Tadro4 and its world. We’ve just set up predation as a great candidate for an additional selection pressure. What we need to know much more about, though, is this tricky business of brains and behavioral complexity. At the very least Tadro4 will have to be a smart prey that is able to eat and, at the same time, avoid being eaten.