Stanford Encyclopedia of Philosophy, 2001

There are two interrelated but distinct programs which go by the name "evolutionary epistemology." One focuses on the development of cognitive mechanisms in animals and humans. This involves a straightforward extension of the biological theory of evolution to those aspects or traits of animals which are the biological substrates of cognitive activity, e. g., their brains, sensory systems, motor systems, etc. The other program attempts to account for the evolution of ideas, scientific theories, epistemic norms and culture in general by using models and metaphors drawn from evolutionary biology. Both programs have their roots in 19th century biology and social philosophy, in the work of Darwin, Spencer, James and others. There have been a number of attempts in the intervening years to develop the programs in detail (see Campbell 1974, Bradie 1986, Cziko 1995). Much of the contemporary work in evolutionary epistemology derives from the work of Konrad Lorenz (1977), Donald Campbell (1974, et. al.), Karl Popper (1972, 1984) and Stephen Toulmin (1967, 1972).

The two programs have been labeled EEM and EET. (Bradie, 1986) EEM is the label for the program which attempts to provide an evolutionary account of the development of cognitive structures. EET is the label for the program which attempts to analyze the development of human knowledge and epistemological norms by appealing to relevant biological considerations. Some of these attempts involve analyzing the growth of human knowledge in terms of selectionist models and metaphors (e. g., Popper 1972, Toulmin 1972, Hull 1988). Others argue for a biological grounding of epistemological norms and methodologies but eschew selectionist models of the growth of human knowledge as such (e. g., Ruse 1986, Rescher 1990).

The EEM and EET programs are interconnected but distinct. A successful EEM selectionist explanation of the development of cognitive brain structures provides no warrant, in itself, for extrapolating such models to understand the development of human knowledge systems. Similarly, endorsing an EET selectionist account of how human knowledge systems grow does not, in itself, warrant concluding that specific or general brain structures involved in cognition are the result of natural selection for enhanced cognitive capacities. The two programs, though similar in design and drawing upon the same models and metaphors, do not stand or fall together.

Biological development involves both ontogenetic and phylogenetic considerations. Thus, the development of specific traits, such as the opposable thumb in humans, can be viewed both from the point of view of the development of that trait in individual organisms (ontogeny) and the development of that trait in the human lineage (phylogeny). The development of knowledge and knowing mechanisms exhibits a parallel distinction. We can consider the growth of an individual’s corpus of knowledge and epistemological norms or his or her brain (ontogeny) or the growth of human knowledge and establishment of epistemological norms across generations or the development of brains in the human lineage (phylogeny). The EEM/EET distinction cuts across this distinction since we may be concerned either with the ontogenetic or phylogenetic development of, e. g., the brain or the ontogenetic or phylogenetic development of norms and knowledge corpora. One might expect that since current orthodoxy maintains that biological processes of ontogenesis proceed differently from the selectionist processes of phylogenesis, evolutionary epistemologies would reflect this difference. Curiously enough, however, for the most part they do not. For example, the theory of "neural Darwinism" as put forth by Edelman (1987) and Changeaux (1985) offers a selectionist account of the ontogenetic development of the neural structures of the brain. Karl Popper’s conjectures and refutations model of the development of human knowledge is a well known example of a selectionist account which has been applied both to the ontogenetic growth of knowledge in individuals as well as the trans-generational (phylogenetic) evolution of scientific knowledge. B. F. Skinner’s theory of operant conditioning, which deals with the ontogenesis of individual behavior, is explicitly based upon the Darwinian selectionist model. (Skinner 1981)

A third distinction concerns descriptive versus prescriptive approaches to epistemology and the growth of human knowledge. Traditionally, epistemology has been construed as a normative project whose aim is to clarify and defend conceptions of knowledge, foundations, evidential warrant and justification. Many have argued that neither the EEM programs nor the EET programs have anything at all to do with epistemology properly (i. e., traditionally) understood. The basis for this contention is that epistemology, properly understood, is a normative discipline, whereas the EEM and EET programs are concerned with the construction of causal and genetic (i.e., descriptive) models of the evolution of cognitive capacities or knowledge systems. No such models, it is alleged, can have anything important to contribute to normative epistemology (e.g., Kim 1988). The force of this complaint depends upon how one construes the relationship between evolutionary epistemology and the tradition.

There are three possible configurations of the relationship between descriptive and traditional epistemologies. (1) Descriptive epistemologies can be construed as competitors to traditional normative epistemologies. On this view, both are trying to address the same concerns and offering competing solutions. Riedl (1984) defends this position. A standard objection to such approaches is that descriptive accounts are not adequate to do justice to the prescriptive elements of normative methodologies. The extent to which an evolutionary approach contributes to the resolution of traditional epistemological and philosophical problems is a function of which approach one adopts (cf. Dretske 1971, Bradie 1986, Ruse 1986, Radnitsky and Bartley 1987, Kim 1988). (2) Descriptive epistemology might be seen as a successor discipline to traditional epistemology. On this reading, descriptive epistemology does not address the questions of traditional epistemology because it deems them irrelevant or unanswerable or uninteresting. Many defenders of naturalized epistemologies fall into this camp (e.g., Munz 1993). (3) Descriptive epistemology might be seen as complementary to traditional epistemology. This appears to be Campbell's view. On this analysis, the function of the evolutionary approach is to provide a descriptive account of knowing mechanisms while leaving the prescriptive aspects of epistemology to more traditional approaches. At best, the evolutionary analyses serve to rule out normative approaches which are either implausible or inconsistent with an evolutionary origin of human understanding.

EEM programs are saddled with the typical uncertainties of phylogenetic reconstructions. Is this or that organ or structure an adaptation and if so, for what? In addition, there are the uncertainties which result from the necessarily sparse fossil record of brain and sensory organ development. The EET programs are even more problematic. While it is plausible enough to think that the evolutionary imprint on our organs of thought influences what and how we do think, it is not at all clear that the influence is direct, significant or detectible. Selectionist epistemologies which endorse a "trial and error" methodology as an appropriate model for understanding scientific change are not analytic consequences of accepting that the brain and other ancillary organs are adaptations which have evolved primarily under the influence of natural selection. The viability of such selectionist models is an empirical question which rests on the development of adequate models. Hull’s (1988) is, as he himself admits, but the first step in that direction. Cziko (1995) is a manifesto urging the development of such models (cf. also the evolutionary game theory modeling approach of Harms 1997). Much hard empirical work needs to be done to sustain this line of research. Non-selectionist evolutionary epistemologies, along the lines of Ruse (1986), face a different range of difficulties. It remains to be shown that any biological considerations are sufficiently restrictive to narrow down the range of potential methodologies in any meaningful way.

Nevertheless, the emergence in the latter quarter of the twentieth century of serious efforts to provide an evolutionary account of human understanding has potentially radical consequences. The application of selectionist models to the development of human knowledge, for example, creates an immediate tension. Standard traditional accounts of the emergence and growth of scientific knowledge see science as a progressive enterprise which, under the appropriate conditions of rational and free inquiry, generates a body of knowledge which progressively converges on the truth. Selectionist models of biological evolution, on the other hand, are generally construed to be non-progressive or, at most, locally so. Rather than generating convergence, biological evolution produces diversity. Popper’s evolutionary epistemology attempts to embrace both but does so uneasily. Kuhn’s "scientific revolutions" account draws tentatively upon a Darwinian model, but when criticized, Kuhn retreated. (cf Kuhn 1972, pp 172f with Lakatos and Musgrave 1970, p. 264) Toulmin (1972) is a noteworthy exception. On his account, concepts of rationality are purely "local" and are themselves subject to evolution. This, in turn, seems to entail the need to abandon any sense of "goal directedness" in scientific inquiry. This is a radical consequence which few have embraced. Pursuing an evolutionary approach to epistemology raises fundamental questions about the concepts of knowledge, truth, realism, justification and rationality.

The interested reader should consult the extensive bibliography originally developed by Donald Campbell and maintained by Gary Cziko at <http://faculty.ed.uiuc.edu/g-cziko/stb/>.

Every scientific enterprise requires formal and semi-formal models which allow the quantitative characterization of its objects of study. The attempt to transform the philosophical study of knowledge into a scientific discipline which approaches knowledge as a biological phenomenon is no different. Much of the evolutionary epistemology literature has been concerned with how to conceive of knowledge as a natural phenomenon, what difference this would make to our understanding of our place in the world, and with answering objections to the project. There are, as well, a number of more technical projects which attempt to provide the theoretical tools necessary for a naturalistic epistemology.

In the simplest sort of model, an organism has to deal with an environment that has two states, S₁ and S₂, and has two possible responses R₁ and R₂. We suppose that what the organism does in each state makes a difference to its fitness. Fitnesses are usually written characterized by a matrix W.

The individual elements of the matrix W_ij are the fitness consequences of response i in state j. So, for instance, W₂₁ denotes the fitness consequences of R₂ in S₁. If we let W₁₁ and W₂₂ equal one and W₁₂ and W₂₁ equal zero, then there is a clear evolutionary advantage to performing R₁ in S₁ and R₂ in S₂.

However, the organism must first detect the state of the environment, and detectors are not in general perfectly reliable. If the organism responds automatically to the detector, we can use the probabilities of responses given states to characterize the reliability of the detector. We write the probability of R₁ given S₁ as Pr(R₁|S₁). This allows us to calculate that responding to the detector rather than always choosing R₁ or R₂ will be advantageous just in case the following inequality holds (cf. Godfrey-Smith 1996):

Pr(R₂|S₂) /(1 Pr(R₁|S₁))   >   Pr(S₁)(W₁₁ W₂₁)/(1Pr(S₁))(W₂₂ W₁₂)

This simple model demonstrates that whether or not flexible responses are adaptive depends on the particular characteristics of the fitness differences that the responses make, the probability of the various states of the environment, and the reliability of the detector. The particular result is calculated assuming that detecting the environmental state and the flexible response system is free in evolutionary terms. More complete analyses would include the costs of these factors.

Static optimization models like the one outlined above can be extended in several ways. Most obviously, the number of environmental states and organismic responses can be increased, but there are other modifications that are more interesting. Signal detection theory, for instance, models the detectors and cues in more detail. In one example, a species of "sea moss" detects the presence of predatory sea slugs via a chemical cue. They respond by growing spines, which is costly. The cue in this case, the water-borne chemical, comes in a variety of concentrations, which indicate various levels of danger. Signal detection theory allows us to calculate the best threshold value of the detector for the growing of spines. 

Static models depict evolutionary processes in terms of fitness costs and benefits. They are static in the sense that they model no actual process, but merely calculate the direction of change for different situations. If fitness is high, a type will increase, if low it will decrease. When fitnesses are equal, population proportions remain at stable equilibrium. Dynamic models typically employ the kinds of calculations involved in static models to depict actual change over time in population proportions. Instead of calculating whether change will occur and in what direction, dynamic models follow change.

Population dynamics, sometimes referred to as "replicator dynamics", offers a tractable way to model the evolution of populations over time under the kinds of selective pressures that can be characterized by static optimization models. This is often necessary, since the dynamics of such populations are often difficult to predict purely on the basis of static considerations of payoff differences. The so-called "replicator dynamics" were named by Taylor and Jonker (1978) and generalized by Schuster and Sigmund (1983) and Hofbauer and Sigmund (1988). They trace their source back to the seminal work of R.A. Fisher in the 1920’s and 30’s. The generalization covers evolutionary models used in population genetics, evolutionary game theory, ecology, and the study of prebiotic evolution. The models can be implemented either mathematically or computationally, and can model either stepwise (discrete) or continuous evolutionary change.

Population dynamics models the evolution of populations. A population is a collection of individuals, which are categorized according to type. The types in genetics are genes, in evolutionary game theory, strategies. The types of interest in epistemological models would be types of cognitive apparatuses, or cognitive strategies -- ways of responding to environmental cues, ways of manipulating representations, and so forth. Roughly, EEM models focus on the inherited and EET models focus on the learned. The evolution of the population consists in changes of the relative frequency of the different types within the population. Selection, typified by differential reproductive success, is represented as follows. Each type has a growth rate or "fitness", designated by w, and a frequency designated by p. The frequency of type i at the next generation p_i is simply the old frequency multiplied by the fitness and divided by the mean fitness of the population "".

p_i   =   p_i w_i /

Division by has the effect of "normalizing" the frequencies, so that they add up to one after each is multiplied by its fitness. It also makes evident that the frequency of a type will increase just in case its fitness is higher than the current population average.

Fitnesses, which should be understood simply as the aggregation of probable-growth factors that drive the dynamics of large populations, may depend on a variety of factors. Fitness components differ from variation components in that they affect population frequencies proportionally to those frequencies, that is to say, multiplicatively. Fitness component in biological evolution include mortality and reproductive rate. In cultural evolution, they include transmission probability and rejection probability. Within either sort of model, what matters is how fitnesses change as a result of other changing factors within the model. In the simplest cases, fitnesses are fixed and the type with the highest fitness inevitably dominates the population. In more complex cases, fitnesses may depend on variable factors like who one plays against, or the state of a variable environment. Most commonly, variable fitnesses are calculated using a payoff matrix like the one above. In general, to calculate the expected fitness of a type, one multiplies the fitness a type would have in each situation times the likelihood that individuals in the population will confront that situation and adds the resulting products.

w_i   =   S_A Pr(A)W_iA

where W_iA is type i’s fitness in situation A. This sort of calculation assumes that the effects of the various situations are additive. More complex situations can be modeled, of course, but additive matrices are the standard. It should be noted, however, that matrix-driven evolution can exhibit quite complex behavior. For instance, chaotic behavior is possible with as few as four strategies (Skyrms 1992).

Some relationships may be represented without a matrix. Boyd and Richerson (1985), for instance, were interested in a special kind of frequency dependent transmission bias in culture, where being common conferred an advantage due to imitators "doing as the Romans do." In such a case, the operative fitness of the type is just the fitness as calculated according to the usual factors, and then modified as a function of the frequency of the type.

The conceptual bases of replicator dynamics are quite straightforward. Getting results typically requires one of two approaches. In order to prove more than rudimentary mathematical results, one typically needs to derive a continuous version of the dynamics. The basic form is

dp_i/dt   =   p(w_i )

with fitnesses calculated as usual. Mathematical approaches have been quite productive, though the bulk of theoretical results apply primarily to population genetics. See Hofbauer and Sigmund (1988) for a compendium of such results, as well as a reasonable graduate-level introduction to the mathematical study of evolutionary processes.

The second approach is computational. With the increase in power of personal computers, computational implementation of evolutionary models become increasingly attractive. They require only rudimentary programming skills, and are in general much more flexible in the assumptions they require. The general strategy is to create an array to hold population frequencies and fitnesses, and then a series of procedures (or methods or functions) which

1. calculate fitnesses,

2. update frequencies with the new fitnesses, and

3. manage interface details like outputting the new state of the population to a file or the screen.

A loop then runs the routines in sequence, over and over again. Most modelers are happy to put their source code on the internet, which is probably the best place to find it.

Part of the difficulty in understanding cognitive behavior as the product of evolution is that there are at least three very different evolutionary processes involved. First, there is the biological evolution of cognitive and perceptual mechanisms via genetic inheritance. Second, there is the cultural evolution of languages and concepts. Third, there is the trial-and-error learning process that occurs during an individual’s lifetime. Moreover, there is some reason to agree with Donald T. Campbell that understanding human knowledge fully will require understanding the interaction between these processes. This requires that we be able to model both processes of biological and cultural evolution. There are by now a number of well-established models of biological evolution. Cultural evolution presents more novelty.

Perhaps the most popular attempt to understand cultural evolution is Richard Dawkins’ (1976) invention of the "meme." Dawkins observed that what lies at the heart of biological evolution is differential reproduction. Evolution in general was then the competitive dynamics of lineages of self-replicating entities. If culture was to evolve, on this view, there had to be cultural "replicators", or entities whose differential replication in culture constituted the cultural evolutionary process. Dawkins dubbed these entities "memes", and they were characterized as informational entities which infect our brains, "leaping from head to head" via what we ordinarily call imitation. Common examples include infectious tunes, and religious ideologies. The main difficulty with this approach has been with providing specifications for the basic entities. The identity conditions of genes can be given, in theory, in terms of sequences of base pairs in chromosomes. There appears to be no such fundamental "alphabet" for the items of cultural transmission. Consequently, the project of "memetics" as contending basis for evolutionary epistemology is on hold pending an adequate understanding of its basic ontology. [See the online Journal of Memetics for more information.]

Population models have been used to good effect in modeling cultural transmission processes. Evolutionary game theory models are frequently claimed to cover both processes in which strategies are inherited and those in which they are imitated. This application is possible in the absence of any specification of the underlying nature of strategies, for instance, whether they are to be thought of as "things" which are replicated, or whether they are properties or states of the individuals whose strategies they are. This is sometimes referred to as the "epidemiological approach", though again, the comparison to infection is due to the quantitative tools used in analysis rather than to any presupposition regarding the underlying ontology of cultural transmission.

The kind of levels involved in evolutionary epistemology are quite different than the kind of levels of selection which are discussed much more often in the "levels of selection" debate in evolutionary biology. In evolutionary biology, the "levels" of selection under discussion are levels of scale. The debate concerns whether genes are always the "units" or "targets" of selection, or whether selection can occur on higher levels, like organisms, groups, and species. The levels involved in evolutionary epistemology, on the other hand, are levels of the regulatory hierarchy involved in the control of behavior. These include the genetic bases of cognitive and perceptual hardware, concepts, languages, techniques, beliefs, preferences, and so forth. Note that in the case of evolutionary epistemology, the terms "levels" and "hierarchy" may be impressionistic. There is often no clear arrangement of levels at all.

There are at least two different approaches that have been taken to modeling multi-level evolution.

1. Dual Transmission Models: Boyd and Richerson (1985) adapted models from genetics to model a case in which a trait (cooperation) was affected both by genetic and cultural evolution. It was first shown that a genetically determined bias on cultural transmission could be selected for in a migratory population. The bias made it easier to pick up local customs, increasing the likelihood of imitation beyond that determined by the frequency and perceived value of the behavior. Once this bias was in place, its effect was strong enough to overcome the perceived costs involved in cooperative behavior. The model yielded two important results. First, it provided a novel mechanism according to which cooperative behavior can stabilize in migratory populations. But more importantly, it demonstrated that cultural evolution cannot be predicted purely on the basis of genetic fitnesses.

2. Multiple Population Models: Harms (1997) constructed a multi-level dynamic population model of bumblebee learning. Mutual information between distributions of sensor types, overt foraging behaviors, and internal foraging preferences on the one hand and environmental states on the other was assessed and compared to average fitness of the population states. It was shown that information present in overt behaviors may be underutilized, and that exaptation of sensor mechanisms for preference formation can bring about the utilization of that information.

Full descriptive accounts of truth and justification both demand a theory of meaning. Until a sign has meaning, it cannot be true or false. Moreover, determining the meaning of justificatory claims may provide a descriptive theory of justification. Presumably, what makes a claim of justification true is the basis of that justification. If meaning is conventional, then the evolution of meaning becomes an instance of the evolution of conventions.

Models of the evolution of conventions have in one case been extended to apply to meaning conventions. Skyrms (1996, chapter 5) gave an evolutionary interpretation of  David Lewis’ (1969) model of rational selection of meaning conventions. Skyrms was able to show that there is strong selection on the formation of "signaling systems" in mixed populations with a full set of coordinated, countercoordinated, and uncoordinated strategies. It is significant that the structure of the model and the selective process by which meaning conventions emerge and are stabilized largely parallels the account of the evolution of meaning given by Ruth Millikan (1984).

In the simplest version, the model is constructed as follows: We imagine that there are two states of affairs T, two acts A, and two signals M. Players have an equal chance of being in either the position of sender, or receiver. Receivers must decide what to do based purely on what the sender tells them. In this purely cooperative version, each player gets one point if the receiver does A₁ if the state is T₁ or A₂ if the state is T₂.

Since players will be both sender and receiver, they must have a strategy for each situation. There are sixteen such strategies, and we suppose them to be either inherited (or learned) from biological parents, or imitated on the basis of perceived success in terms of points earned. Strategies I₁ and I₂ are signaling systems, in that if both players play the same one of these two strategies they will always get their payoff. I₃ and I₄ are anti-signaling strategies, which result in consistent miscoordination, though they do well against each other. All of the other strategies involve S₃, S₄, R₃, or R₄, which results in the same act being performed no matter what the external state is.

Bibliography