A new defense of adaptationism

A new defense of adaptationism

Adaptive explanations are the bread and butter of evolutionary biology. But the scientific legitimacy of such adaptive explanations is controversial, largely because of the classic paper “The Spandrals of San Marco and the Panglossian Paradigm: A Critique of the Adaptationist Programme” published by Stephen Jay Gould and Richard Lewontin in 1979. The controversy persists to this day in large part, I believe, because the fundamental challenge raised by Gould and Lewontin has not yet been met; in fact, it is rarely even acknowledged. The objectification of value in teleology described above can change this status quo, for we can now to defend the scientific legitimacy of adaptive explanations in a new and deeper way.

First, some terminology. I will refer to claims to the effect that a trait is an adaptation as an adaptive hypothesis. A specific adaptive hypothesis is a claim to the effect that a trait is an adaptation for some specified adaptive function, and a general adaptive hypothesis claims that a trait is an adaptation but identifies no adaptive function. A general adaptive hypothesis expresses the presupposition that the trait has some adaptive explanation. An example is the claim that large primate testes are an adaptation. An example of a specific adaptive hypothesis is the claim that large primate testes are an adaptation for producing more sperm. An adaptive explanation of a trait explains its existence or persistence as a result of adaptive evolution, i.e., by means of natural selection for that trait. Finally, by adaptationism I mean the thesis that the activity of pursuing adaptive explanations of the existence and nature of biological traits is a normal and legitimate part of empirical science.¹²

Gould and Lewontin central complaint about adaptive explanations is that we have no principled way to tell when they are needed. People deploy adaptive explanations without justifying them over non-adaptive alternatives, such as appeals to architectural constraints or genetic drift. If one adaptive explanation fails it is simply replaced by another, but sufficient ingenuity enables any trait to be given an adaptive explanation. The general adaptive hypothesis that a trait is an adaptation is treated as untestable. As Lewontin puts it, “the adaptationist program makes of adaptation a metaphysical postulate that … cannot be refuted” because the presupposition that a trait is an adaptation is never questioned (Lewontin 1977/1985, p. 76). The deeper worry is that the presupposition that a trait is an adaptation is really is untestable.¹³ There is a thicket of alternatives to adaptive explanations. How in principle can we tell when it is appropriate to pursue adaptationist branches? Gould and Lewontin summarize the predicament thus:

We would not object so strenuously to the adaptationist programme if its invocation, in any particular case, could lead in principle to its refutation for want of evidence. We might still view it as restrictive and object to its status as an argument of first choice. But if it could be dismissed after failing some explicit test, then alternative would get their chance. (1979, pp. 258f).

The fundamental challenge, then, is to find some empirical test for general adaptive hypotheses. Without such a test, how can the practice of giving adaptive explanations be a normal and legitimate part of empirical science? In other words, the thesis of adaptationism would seem to be false.

Gould and Lewontin’s challenge to adaptationism provoked a storm of response. So many of the responses share the same basic form that this form can be called the “canonical” response. In a nutshell, the canonical response is to concede that there is no general empirical test for general adaptive hypotheses but construe this as on a par with normal empirical science.

Richard Dawkins nicely illustrates the cannonical response when he considers traits that might not be adaptations. He points out that it is possible to test rival adaptive hypotheses by ordinary scientific methods, noting that “hypotheses about adaptation have shown themselves in practice, over and over again, to be easily testable, by ordinary, mundane methods of science” (Dawkins, 1983b, pp. 360f). Dawkins’s central point is that specific adaptive hypotheses have observable consequences, so they entail empirical predictions and thus can be tested. Dawkins illustrates this point with primate testes size. As it happens, primate testes size scales roughly but not exactly with body size. If testes weight is plotted against body weight, there is considerable scatter around the average line.

A specific adaptive hypothesis is that in those species in which females mate with more than one male, the males need bigger testes than in those species in which mating is monogamous or polyganous: A male whose sperms may be directly competing with the sperms of another male in the body of a female needs lots of sperms to succeed in the competition, and hence big testes. Sure enough, if the points on the testis-weight/body-weight scattergram are examined, it turns out that those above the average line are nearly all from species in which females mate with more than one male; those below the line are all from monogamous or polygynous species. The prediction from the adaptive hypothesis could easily have been falsified. In fact it was borne out… (1983b, p. 361)

This illustrates how specific adaptive hypotheses can be tested by ordinary empirical methods.

Note, though, that Dawkins does not address the testability of general adaptive hypotheses. Furthermore, the test for specific adaptive hypotheses cannot be used to produce a test general adaptive hypotheses. The observable consequences of a specific adaptive hypothesis depend on the specific function hypothesized. Different functions may well entail different predictions. For example, the hypothesis that large primate testes are an adaptation for temperature regulation would entail a quite different prediction about where species fall in the testis-weight/body-weight scattergram. By contrast, the general hypothesis that large testes are an adaptation for something or other entails no prediction about where species fall in the scattergram. So, a general adaptive hypothesis inherits no observational consequences from specific hypotheses. For this reason, Dawkins admits that general adaptive hypotheses are untestable. “It is true that the one hypothesis that we shall never test is the hypothesis of no adaptive function at all, but only because that is the one hypothesis in this whole area that really is untestable” (1983b, p. 361). In other words, Dawkins thinks the fundamental challenge to adaptationism cannot be met. Of course, evidence for a specific function is a fortiori evidence for some function, so corroborating a specific adaptive hypothesis also corroborates the corresponding general hypothesis. But we cannot test all possible specific adaptive hypothesis for a trait. So the testability of specific hypotheses provides no test for general hypothesis.

The cannonical response is weak.¹⁴ It capitulates in the face of Gould’s and Lewontin’s fundamental challenge by agreeing that there is no test for general adaptive hypotheses. But plenty of traits are not adaptations, and adaptive explanations are often inappropriate. Is there really no empirical way to tell whether adaptive explanations are in the offing? I think the answer is “Yes” and the objectification of biological teleology provides the key. The test we need is simply to collect and analyze evolutionary activity information.

The sign that an evolutionary process is creating adaptations is that its activity data are significantly higher than what would be expected if selection were random. If activity waves rise above the noise generated in a no-adaptation neutral model, then you know the corresponding items are adaptations even if you are ignorant about the adaptive functions. The activity data show that some adaptive explanation is needed even if it silent about the merits of any specific explanation.¹⁵ In other words, the evolutionary activity method tests general rather than specific adaptive hypotheses.

Thus, the evolutionary activity test directly responds to the fundamental challenge to adaptationism. It parts company with the canonical response by not capitulating to Gould and Lewontin. As far as I know, it is the first response that takes this bull by the horns. The test does not assume that traits are adaptations but tests whether they are. Adaptive “just-so” stories have no place here; such stories propose specific adaptive hypotheses and these are not at issue. The issue is general adaptive hypotheses, and these are accepted skeptically. Where the canonical response is weak, the activity test is strong. It makes the question of adaptation objective and empirical. When the adaptive stance is adopted, it is on the basis of empirical evidence against nonadaptive alternatives. So we can pursue the adaptationist program constructively and self-critically, as a normal and legitimate part of empirical science. Gould and Lewontin said that they “would not object to strenuously to the adaptationist programme if its invocation, in any particular case, could lead in principle to its refutation for want of evidence” (1979, pp. 258f). The evolutionary activity method provides just the sort of tool that Gould and Lewontin sought. So if we can take them at their word, they should now withdraw their objection.

Towards a comparison on biological and cultural evolution

The information in evolutionary activity graphs can be summarized with statistics that reflect how evolution is creating adaptations. Such statistics have various uses, such as enabling quantitative comparisons of adaptive evolution across different systems. After informally explaining some of these statistics and explaining how they have been used to classify evolving systems, I will show how they shed new light on the relationship between biological and cultural evolution.¹⁶

When attempting to measure the degree of adaptive evolution in a system, one might try to reflect at least three different things. First, one might try to capture how well adapted the adaptations are, that is, how optimally they perform their function. Second, one might try to reflect the intensity of adaptive evolution, that is, the rate at which new adaptations are being produced by natural selection. Third, one might try to reflect the extent of adaptive evolution, that is, the total continual adaptive success of all the adaptations in the system. I will concentrate on the second and third ideas.

The intensity of evolutionary activity intuitively corresponds to the rate at new evolutionary activity is being created in the system, measured as the rate at which new activity waves are entering the activity graph. When there are very few new waves the intensity of activity is low; when a lot of new waves are being generated the intensity is high. To clean up this measure of the intensity of evolutionary activity, one would normalize the intensity observed in the target system with the intensity observed in a neutral model, yielding the excess intensity. One simple way to accomplish this is to measure not new waves but new waves that accrue more activity than would be expected in a neutral model.

The extent of evolutionary activity intuitively corresponds to the amount of evolutionary activity present in the system, measured as the sum total of the activity in an activity graph at a given time. If you think of the activity waves as being made up of grains of sand, the extent of activity at a given time is the mass of sand at that time in the graph, where the mass is weighted by its height in the graph. When the system has lots of very large activity waves, the extent of activity is very high. When the system has only a few waves and they are relatively low, the extent of activity is relatively low. As with intensity, one would clean up this measurement by normalizing the extent of activity observed in the target system with the extent observed in a neutral model, perhaps simply by subtracting the neutral extent from the target extent, thus yielding the excess extent.

The extent and intensity of evolutionary activity are two independently varying aspects of a system’s adaptive evolution. For example, if adaptations continue to persist indefinitely without changing and no new adaptive innovations invade the system, then the extent of activity will continually increase, but the intensity of activity will fall to nil. On the other hand, if evolution is continually creating new adaptations and destroying older ones, the intensity of activity will be positive but the extent of activity will be very low.

The intensity and extent of activity statistics are quite general and apply to data generated by both artificial and natural systems, and they apply at different levels of analysis. I have used evolutionary activity statistics to measure the creation of adaptations in a variety of evolutionary system (Bedau and Packard 1992; Bedau 1995; Bedau 1996; Bedau and Brown 1997; Bedau, Joshi, and Lillie 1999; Bedau, Snyder, Brown, and Packard 1997; Bedau, Snyder, and Packard 1998; Rechtsteiner and Bedau 1999a,b). Comparing data from a variety of different systems suggests that these statistics can be used to partition evolutionary dynamics into four qualitatively different classes. Class 1 consists of systems in which evolution creates no adaptations at all (e.g., all neutral models, systems in which the mutation rate is too high, and systems in which the selection pressure is too low, etc.). The signature for this class is zero excess intensity and extent of activity. Systems in which evolution has created adaptations but in which no new adaptations are being created fall into class 2 (e.g., stable ecosystems), with the signature of zero excess intensity and unbounded excess extent. Class 3 consists of systems that continually create new adaptations but are bounded in the amount of adaptive structure they contain (e.g., if new adaptations always supplant old adaptations). It’s signature is positive excess intensity and bounded excess extent. If new adaptations are continually created and the total amount of adaptive structure continues to grow, then the system falls into class 4, which has the signature of positive excess intensity and unbounded excess extent. The biosphere as reflected in the fossil record exhibits class 4 dynamics. (For more details about this classification, see Bedau, Synder, and Packard 1998, and Skusa and Bedau 2002.)

Class 4 is an especially explosive kind of evolutionary creativity. It is intriguing in part because no known existing artificial evolving system generates the same kind of behavior (Bedau, Snyder, Brown, and Packard 1997; Bedau, Snyder, and Packard 1998). Although we do not know the mechanism behind class 4 behavior, but it seems to involve the course of evolution continually creating new kinds of environments that open the door to qualitatively new kinds of adaptations. There is some reason to think that a similar hyper-creative process might be at work in cultural evolution. We could start to assess this conjecture if we could apply evolutionary activity statistics to cultural evolution. I have recently started to do this in collaboration with Andre Skusa. Specifically, we have examined the evolution of technology as reflected in patent records, and we use evolutionary activity to create an empirical picture of the adaptive dynamics in patented inventions. Such pictures allow us to compare the dynamics of patented technology with those exhibited in biological evolution.¹⁷

Patents offer some important advantages for those looking for cultural evolution in empirical data. It is often difficult to operationalize the units of cultural evolution. It is difficult to distinguish new innovations from copies of old innovations when the items are ideas or other mental aspects of culture. Another difficulty is ascertaining precise genealogical relationships. One can finesse these difficulties by studying the evolution of technology as reflected in patent records. Although the evolution of inventions involves the diffusion and selection of ideas, one can identify individual inventions with individual patents. To be patentable an invention must meet three criteria: novelty, usefulness, and non-obviousness. So patented inventions are certified to be new and functional. A patent's novelty is documented by citing the previous patents (and sometimes published papers) that involve related ideas; these are called the patent's “prior art.” The citations should identify all the important prior art from which the invention is derived, and in the aggregate they allow a patent’s entire genealogy to be inferred.

The analogies and disanalogies between biological and cultural evolution are a matter of some controversy (Hull 1988, 2001) but it is relatively straightforward to extract evolutionary activity data from patent records. The units of evolution with which we are concerned (at least in the first instance) are individual patents; these are analogous to genes (or, as memeticists might suggest, “memes”). A gene could vanish forever from an evolutionary system. By contrast, a patented invention never goes fully extinct because the invention exists forever in the patent records. We consider that a patent “reproduces” when it leads to the production of other patents; that is, in contrast with most biological evolution, patent reproduction necessarily involves evolutionary innovation.

Especially successful or valuable patents tend to be those that are especially heavily cited. A large body of work in scientometrics has repeatedly confirmed that number of citations is a good reflection of the technological significance and economic value of a patented invention (Albert et al. 1991; Narin 1994; Pavitt 1985; Perko and Narin 1997; Albert 1998). Once a patent has received more than ten citations, the economic value reflected by each additional citation has been estimated to be more than one million US dollars (Harhoff et al. 1999). For these reasons our bookkeeping of an individual patent's evolutionary activity is based on summing the citations the patent has received. From this perspective, the adaptive success of a patented innovation is measured by the extent to which it spawns subsequent patented innovations. More specifically, we increment a patent's activity at a given time by the number of citations it receives from patents issued at that time.

The fact that a patent has received a few citations does not prove that the invention significantly shapes the evolution of subsequent inventions. A patent might be cited by one or two subsequent patents even if patents to cite were chosen entirely at random. As with evolutionary activity measurements in other contexts, we can evaluate a patent's adaptive significance by comparing its activity with the activity observed in a neutral model of patent evolution. Our patent neutral model mirrors a few key aspects of the real patent data. In both the same number of patents are issued each week, and they exhibit the same distribution into the various patent classes. Patent citations refer to the same number of pre-9/96 and post-9/96 patents, and the references to post-9/96 patents fall into the various patent classes according to the same distribution. The key distinguishing feature of the neutral model is that the patents to be cited are always chosen randomly.

Note that the activity of one patent in Figure 4 stands far above the rest, accruing almost twice as many citations as any other patent. This patent covers the technology that allows web browsers to display information such as advertisements while a page is being loaded a link is clicked. The second most heavily cited patent covers the technology that allows cell phones to receive email and faxes, and the third most heavily cited patent allows remote control of the receipt and delivery of wireless and wireline voice and text messages. All of the ten most heavily cited patents fall into the information technology sector, and seven of them involve the Internet.

More detailed information can be extracted from the evolutionary activity data (see Skusa and Bedau 2002). The point here is simply that evolutionary activity statistics make it feasible to visualize and quantitatively assess the adaptive evolutionary dynamics exhibited in cultural evolution. We have applied the method to technological evolution as reflected in patent record data, but it can be applied to a variety of other cultural systems. Our pilot project underscores the vast importance of information technology, and especially the Internet, over the past five years. This is not news, of course; it just corroborates what we already knew. But it does confirm the aptness and probity of evolutionary activity analysis of cultural evolution. Furthermore, it opens the door to quantitative comparison of cultural and biological evolution. And this provides a constructive empirical route for investigating whether the hyper-creativity exhibited by biological evolution also characterizes cultural change.