On Meme-Gene Coevolution

Larry Bull, Owen Holland# ,* and Susan Blackmore+

Faculty of Computer Studies and Mathematics

#Faculty of Engineering

+Faculty of Applied Sciences

University of the West of England

Bristol BS16 1QY, U.K.

larry@[Link]

*Microsystems Laboratory

Department of Electrical Engineering

California Institute of Technology (136-93)

Pasadena, CA 91125, USA

Abstract

In this paper we examine the effects of the emergence of a new replicator, memes, on the evolution of a pre-existing

replicator, genes. Using a version of the NKCS model we examine the effects of increasing the rate of meme

evolution in relation to the rate of gene evolution, for various degrees of interdependence between the two

replicators. That is, the effects of memes’ (suggested) more rapid rate of evolution in comparison to that of genes

is investigated using a tunable model of coevolution. It is found that, for almost any degree of interdependence

between the two replicators, as the rate of meme evolution increases, a phase transition-like dynamic occurs under

which memes have a significantly detrimental effect on the evolution of genes, quickly resulting in the cessation

of effective gene evolution. Conversely, the memes experience a sharp increase in benefit from increasing their

rate of evolution. We then examine the effects of enabling genes to reduce the percentage of gene-detrimental

evolutionary steps taken by memes. Here a critical region emerges as the comparative rate of meme evolution

increases, such that if genes cannot effectively select memes a high percentage of the time, they suffer from meme

evolution as if they had almost no selective capability.

1
1. Introduction

Dawkins [1976] coined the term "meme" to describe information copied from person to person by imitation.

Memes vary, are selected, and are inherited (i.e. copied by imitation); they therefore fit the evolutionary algorithm

and can be considered as a replicator [Dennett 1995]. Most theories of human evolution, including sociobiological

theories, assume that natural selection has operated on only one replicator, the genes. However, humans are adept

at imitation almost from birth [Meltzoff 1996] and therefore live in an environment pervaded by memes. We should

therefore consider human evolution to be a product of two replicators - memes and genes.

Blackmore [1999a] has argued that competition between these replicators is responsible for both the evolution of

language and the enormous increase in brain size in humans. In a process termed "memetic driving" the most

successful memes force the genes to create machinery that is especially good at copying those memes. The process

works as follows: competition between memes results in some memes becoming more successful and being widely

imitated, while others fail. These successful memes will include those that have been copied because they are

useful (such as new skills and technologies) and some that are useless or even harmful. People who are best at

copying the most successful memes acquire advantages in terms of status and survival, as well as being

preferentially selected as mates. Therefore genes that facilitate the imitation of those particular memes are favoured

and increase in the gene pool. One simple application of this general argument is to brain size. If we assume that

having a larger brain makes imitation easier, memetic driving will favour genes for larger brains. On this theory

the human brain expanded rapidly in size for memetic, not purely genetic, reasons.

The same process can account for such uniquely human abilities as the enjoyment of music. As soon as early

humans began imitating sounds, memes based on tunes and rythms could spread. Some of these sounds would

evoke emotional reactions, act as signals of status, or have other effects that improved their chances of being

copied. People who could copy the most successful sounds would be at an advantage and so the genes that made

them good at copying those sounds would spread. Musical ability would generally increase in the population,

allowing more memes to spread and increasing the speed of memetic evolution. In this way our musical abilities

have been shaped by a long history of coevolution between the memes that were copied and the machinery that

copied them.

2
As the process continues, the memes evolve faster and faster, outstripping the genes’ ability to change. When this

happens we should expect traits that are adaptive to the memes but maladaptive to the genes to spread. In this

situation the genes would benefit if they could select out maladaptive memes. We arrive at a conception of the

human brain as evolved not only to spread memes, but to select which memes to replicate. In this sense it is

comparable with the immune system, being evolved to select between desirable and undesirable replicators

[Blackmore 1999b].

There have been several previous coevolutionary theories (e.g. [Durham 1991], [Lumsden & Wilson 1981]).

Cavalli-Sforza and Feldman [1981] developed a mathematical model of gene-culture evolution but, in spite of

showing how maladaptive traits could survive (that is, maladaptive to the genes), they ultimately assumed, as

Lumsden and Wilson put it, that "the genes hold culture on a leash". Boyd and Richerson [1985] developed a

mathematical model to explore the conditions under which social learning (including imitation) would be favoured,

and have shown that dual-replicator models can account for group selection processes [1990] but they have not

considered memetic driving. Best [in press] has used a simple simulation environment in which evolving agents

learn by imitation, and shown that imitation is (like individual learning) able to guide genetic evolution towards an

adaptive goal. However, when genes and culture pursued diametrically opposed goals, he found that culture was

no match for genetic evolution with individual learning.

None of these models considers two issues that are important to Blackmore’s coevolutionary model; the relative

speed of the two replicators, and whether one replicator can exert an explicit selection pressure on the other. These

issues are addressed here using versions of the NKCS model [Kauffman 1993].

We examine the effects of different rates of evolution between the two replicators. With low dependence between

the memes and genes, it is found that varying the rate of meme evolution with respect to the rate of gene evolution

has no significant effect on either replicator. That is, both are able to evolve successfully within their respective

attribute spaces. However, with a slightly larger degree of interdependence, increasing the rate of meme evolution

generates a phase transition-like phenomenon under which meme evolution (rapidly) benefits significantly and

gene evolution degrades to a random walk; for any significant amount of inter-replicator dependence, large

benefits/losses can be experienced by altering the difference in evolution rates between them.

3
We then extend the model to examine the effects of genes actively selecting memes such that meme evolution is

restricted, by varying degrees, in the extent to which it may have detrimental effects on the genes. It is found that,

for significant meme-gene interdependence, the benefits to genes of increasing their selective pressure over memes

increases. It is further shown that, as the relative rate of meme evolution increases, significant benefits to genes are

only obtained under high selective pressure.

The paper is arranged as follows: the next section describes the NKCS model. Section 3 presents results from its

use to examine the different relative rates of meme and gene evolution. Section 4 introduces and examines the

version of the model in which genes exert some explicit control over meme evolution. Finally, our findings are

discussed.

2. The NKCS Model

Kauffman [1993] introduced the NKCS model to allow the systematic study of various aspects of natural evolution

between interacting species. However, the model is abstract, and so can be applied to the study of any collection of

interacting replicators, e.g. memes and genes. In the basic model a given population is represented by a single set

of N replicators, each with two possible states, i.e. a population is assumed to be converged to all individuals having

the same value for each constituent replicator. The fitness of an individual depends upon the contribution made by

each replicator, each of which depends upon K other replicators (random) within the individual (epistasis).

Increasing K, with respect to N, increases the epistatic linkage, increasing the ruggedness of the fitness landscape

of a population by increasing the number of fitness peaks, increasing the steepness of the sides of the peaks and

4
decreasing their typical heights. That is, as the dependence between an individual’s aspects increases, the attribute

(a) state fitness

(b)
000 0.32
A 1 1 0 B 0 1 1 001 0.41
010 0.52
011 0.29
N=3 K=1 C=1 S=1 100 0.75
N
(c) 1 101 0.47
Fit = fitness (n) 110 0.36
N
111 0.58
n=1

Figure 1: Showing an example NKCS model. (a) shows each replicator depends on one other locally and on
one in the other population. Therefore there are eight possible configurations, each of which is assigned
a random fitness as shown in (b). Each replicator of each population has such a table created for it. Total
fitness of a given (converged) population of replicators is the averaged sum of these values (c).

space becomes increasingly complex. Each individual replicator is also said to depend upon C replicators (random)

in the other populations with which it interacts. Hence, the adaptive moves made by one population of replicators

may alter the fitness landscapes of its (S) partners; altering C, with respect to N, changes the extent to which

adaptive moves by each individual deforms the landscapes of its partnering populations. As C increases, mean

evolutionary fitness drops and the time taken to reach an equilibrium point increases, where the fitness level of the

equilibrium decreases.

The model assumes all inter-population (C) and intra-population (K) interactions are so complex that it is only

appropriate to assign random values (Gaussian distribution) to their effects on fitness. Therefore for each of the

possible K+C interactions, for each given replicator, a table of 2(K+C+1) fitness contributions is created, with all

entries in the range 0.0 to 1.0, such that there is one fitness for each possible combination of replicators. The fitness

contribution of each replicator within a population is found from its individual table. These contributions are then

summed and normalised by N to give the actual fitness of the individual (Figure 1).

Kauffman considered converged populations of species and random hill-climbing to evolve each in turn, i.e. each

species/population uses the current context of the others to determine whether a random alteration to its

configuration represents progress. That is, from a given configuration, a population randomly alters one replicator’s

state and calculates the resulting fitness. If the new fitness is greater than the population’s current fitness, in the

current environment, the population adopts the new configuration (see [Kauffman 1993] for full details of the

5
model).

3. An NKCS Model of Meme-Gene Coevolution

In this paper we consider the coevolutionary progress of two populations of replicators - memes and genes (S=1).

Each population consists of individuals of sixty-four replicators (N=64), with varying degrees of inter- and intra-

population dependence. We also introduce a new parameter, T, which determines how many generations of meme

evolution occur per generation of gene evolution. In this way we are able to examine the effects of the emergence

of memes upon genes (low T) and the effects of increasing their respective rate (> T). Thus in the standard NKCS

model it is always the case that T=1.

Results were obtained by generating a random landscape for the given set of parameters (K, C, T) and recording

the resultant fitness of each population after five thousand gene generations. For each generated landscape, ten runs

were carried-out, such that the results are the averages of a hundred runs per set of parameters. Note that in the

figure negative values for T imply reciprocal meme rates, e.g. T=-5 means five generations of gene evolution per

generation of meme evolution.

Figure 2 presents the initial results and shows that, for all K, once there is significant dependence between the

memes and genes (roughly C>2), memes can benefit significantly from increasing their rate of evolution with

respect to that of genes. Conversely, increasing the rate of meme evolution can have significantly detrimental

effects on gene evolution. It can be seen that at around T= -30, i.e. thirty generations of gene evolution per meme

generation, the memes begin to have a detrimental effect on the genes. This effect increasingly degrades gene

evolution until around T=10, by which time gene evolution effectively stops, with higher C genes performing about

6
as well as a random walk (fitness = 0.5). However, memes do not experience any significant benefits from an

Figure 2: Showing the effects of varying the rate of meme evolution with respect to gene evolution, for
varying degrees of dependence within and between individuals.

increase in their rate of evolution until around T=-10. The benefits then increase until around T=30, after which no

further benefits are seen. That is, a sharp transition occurs across a range of forty rates which results in the effective

cessation of gene evolution and greatly improved meme evolution. Note that the size of the benefit/detriment in

fitness increases with increasing C.

As stated above, as C increases, the amount of epistatic coupling between the fitness landscapes of the populations

7
of replicators increases. The effect of this coupling is to cause movement in the partners’ landscapes, which

increases the difficulty experienced by evolution in locating high-fitness optima. By increasing their rate of

evolution, memes are able to reduce the effects on themselves of this fitness landscape movement since the increase

in rate reduces the number of possible evolutionary advances/changes which can be made by the genes before the

memes have to "re-adapt" to them. That is, the greater T, the more (temporarily) stationary the memes’ fitness

landscape appears to them. Similarly, for the genes, as the memes are able to make more and more evolutionary

advances before the genes have the chance to re-adapt, the more the genes experience the deleterious effects of the

coupling between the landscapes.

4. Brains as Immune Systems

It has been noted that, since memes have the potential to be maladaptive to genetic evolution, they can be viewed

as analogous to viruses (e.g. [Dawkins 1993][Brodie 1996] - see [Bura 1994] for simulations in which lethal memes

emerge and persist). In the above model the genes had no explicit control over meme evolution; memes were seen

as completely autonomous replicators. However it has recently been suggested that, to some degree, genes may

actively "select" memes, stopping them from evolving in gene-detrimental directions [Blackmore 1999b]. This can

be seen as akin to the situation in which a set of genes’ immune system must attempt to respond to the presence of

a fast evolving virus such as HIV.

We have extended the model described in the previous section to allow genes to explicitly exert a selective pressure

over the memes’ evolutionary process. Now, with a probability represented by a percentage rate β, each time the

meme population finds a new advantageous configuration via mutation, it does not move to that configuration if it

is detrimental to the current configuration of the gene population. Hence, in the previous model the memes were

always able to move to a meme-beneficial configuration (β = 0). Using this model we are able to examine the

effects of altering the number of gene-harmful meme mutations which propagate, i.e. the effects of varying the

genes’ ability to successfully discriminate between memes.

Results were obtained by using the same random landscapes from the previous section, with given sets of

parameters (K, C, T), recording the resultant fitness of each population after five thousand gene generations, for

8
varying β. Again, ten runs on each of the landscapes were carried out such that the results are the averages of a

hundred runs per set of parameters. Note that we only consider positive values of T here, but the same trends are

found for lower rates of meme evolution (not shown).

Figure 3 shows example results for a given set of parameters K,C, and various T, for three values of β. It can be

seen that as the rate of successful gene discrimination increases (>β), i.e. the greater the amount of explicit meme

Figure 3: Showing the effects of varying the rate of meme selection by the genes, for varying rates of meme
evolution. β = 0 corresponds to the case shown in Figure 2.

selection by the genes, the greater the benefits to the genes. At the highest rates the effects found in the previous

version of the model are lost and the genes evolve to fitness levels similar to those found for very low C. Conversely,

the memes do less well as β increases, as expected. This has been found to be the case whenever there is significant

dependence between the replicators. Figure 4 confirms this result, showing the resultant fitnesses after five

thousand gene generations for the range of possible β, and given values of K, C and two values of T.

It is clear that, when there is high dependence between memes and genes, more meme mutations will be produced

which will have an effect on the genes [Kauffman 1993, p.250]. Thus a greater number of gene-detrimental

mutations can arise, and hence the genes will increasingly benefit from restricting meme evolution. As the relative

rate of meme evolution increases, the number of meme mutations increases, again increasing the probability of

gene-detrimental mutations occurring; increasing T is akin to increasing C here. However, meme evolution

effectively stops under high β since fewer and fewer evolutionary steps become possible.

From Figure 4 it can also be seen that, as the relative rate of meme evolution increases, the benefits to genes of

9
increasing β are lower for medium β. Figure 5 shows this result explicitly by comparing the results from runs with

three values of T over a range of β. It can be seen that, whilst in all cases being able to successfully discriminate

between memes a high percentage of the time is very beneficial to genes, this benefit increases more sharply with

Figure 4: Showing the effects of varying the rate of meme selection, for given K, C and T.

increasing β under high T; for high T the benefits of high β decrease more rapidly than for low T. Note also that the

rate of change in this behaviour does not alter linearly since there is a far greater change between T=1 and T=19,

than between T=19 and T=39. Thus as the relative rate of meme evolution increases, genes being able to exert an

explicit selection pressure on meme evolution over seventy percent of the time gives the genes a sustained benefit.

5. Conclusions

10
For some time it has been recognised that, within complex social environments, particularly those involving

humans, a form of cultural replicator exists. Such a replicator must therefore coevolve with genes, the existing

replicators. This coevolution will be influenced by the degree of coupling between the replicators and also by their

relative rates of replication.

In this paper we have presented results from a version of the NKCS model of meme-gene coevolution. It has been

found that, for most degrees of dependence between the two replicators, regardless of the dependence within the

populations, a phase transition-like dynamic occurs as the relative rate of replication is varied. Within our model,

Figure 5: Showing the effects of varying the relative rate of meme evolution on the genes ability to select
over meme mutations.

until the rate of meme evolution is a thirtieth that of genes, genes remain unaffected by their presence. From then

on, until the memes evolve ten times faster than the genes, the genes experience increasingly negative effects from

the presence of the memes, and thereafter are unable to evolve effectively. Conversely, the memes do not

experience any benefit from increasing their rate of evolution until it is around a tenth that of the genes. From then

on, until they evolve thirty times faster than the genes, they experience increasing benefit from increasing their rate

of evolution. Thereafter they suffer no beneficial or detrimental effects from any increase. The symmetry here is

due to the symmetry in the rest of the model. Varying N, e.g. N=128, was found to have little effect (not shown).

The effects of genes being able to exert a selective pressure, of varying degrees, over memes have also been

examined within the model. It has been found that, under conditions of significant meme-gene interdependence, a

critical amount of pressure exists with increasing meme replication rate in which genes experience large benefits,

11
and meme evolution falters. If it is assumed that for larger amounts of selection pressure to be exerted by genes,

bigger brains are required [Blackmore 1999b], then this result adds to a memetic explanation of the origin of

humans’ unexpectedly large brains [Blackmore 1999a].

The concept of phenotypic plasticity, i.e. gene-based lifetime learning of phenotypic traits, whether individual or

through imitation - the Baldwin effect [Baldwin 1896] - has not been examined here. That is, in this model, genes

are not able to respond structurally during their life-time/evaluation to the existence of memes capable of faster

evolution than themselves (T>1). We suggest that such phenotypic plasticity has the potential to be of benefit to

genes, but note that results here show genes suffer most significantly from the presence of memes whilst their rate

of evolution is still greater than that of memes, i.e. before lifetime learning could have any significant effect. Once

the rate of meme evolution is greater than that of genes, phenotypic plasticity could perhaps allow genes to track

meme evolution, such that all effects of the difference in rates would be lost (assuming equivalent effective rates

of learning and meme evolution). However, it is known that too much learning can disrupt the Baldwin effect due

to the reduction in selection pressure [Maynard Smith 1993]; it is therefore possible that for faster rates of meme

evolution, equivalent life time phenotypic plasticity may in itself have detrimental effects on gene evolution. Genes

which further evolve imitation to keep track of rapidly evolving memes have the potential to do themselves more

harm in the long run than if they used individual learning; genes producing learning by imitation not only creates

the necessary conditions for memes to emerge, which may have detrimental effects on gene evolution, but its

continued use has the potential to makes things worse ! We are currently extending the models presented here to

include phenotypic plasticity.

References

12
Baldwin, J. M. (1896) A New Factor in Evolution. American Naturalist 30: 441-451.

Best, M. (in press) How Culture Can Guide Evolution. Adaptive Behaviour.

Blackmore, S. (1999a) The Meme Machine. Oxford University Press.

Blackmore, S. (1999b) Can the Memes Get Off the Leash ? Presented at Do Memes Account for Culture ? King’s

College Research Centre Conference, King’s College, Cambridge 4th-5th July 1999.

Boyd, R. & Richerson, P.J. (1985) Culture and the Evolutionary Process. University of Chicago Press.

Boyd, R. & Richerson, P.J. (1990) Group Selection Among Alternative Evolutionary Stable Strategies. Journal of

Theoretical Biology, 145:331-342.

Brodie, R. (1996) Virus of the Mind: The New Science of the Meme. Integral Press.

Bura, S. (1994) MINIMEME: Of Life and Death in the Nosphere. In D. Cliff, P. Husbands, J-A. Meyer & S. Wilson

(eds.) From Animls to Animats 3. MIT Press, pp479-486.

Cavalli-Sforza, L.L. & Feldman, M.W. (1981) Cultural Transmission and Evolution: A Quantative Approach.

Princeton University Approach.

Dawkins, R. (1976) The Selfish Gene. Oxford University Press.

Dawkins, R. (1993) Mind Viruses. In B. Dahlbohm (ed.) Dennet and his Critics: Demystifying Mind, Blackwell,

pp13-27.

Dennett, D. (1995) Darwin’s Dangerous Idea. Penguin.

13
Durham, W.H. (1991) Coevolution: Genes, Culture and Human Diversity. Stanford University Press.

Kauffman, S. A. (1993) The Origins of Order. Oxford University Press.

Lumsden, C.J & Wilson, E.O. (1981) Genes, Mind and Culture. Harvard University Press.

Maynard Smith, J. (1993), The Theory of Evolution, Cambridge University Press.

Meltzoff, A.N. (1996) The Human Infant as Imitative Generalist: A 20-year progress report on infant imtation with

implications for comparative psychology. In C.M. Heyes & B.G. Galef (eds) Social Learning in Animals: The

Roots of Culture. Academic Press, pp347-370.

14