Biology and Philosophy 16: 655683, 2001.

2001 Kluwer Academic Publishers. Printed in the Netherlands.

Complexity, Self-Organization and Selection

ROBERT C. RICHARDSON
Department of Philosophy
University of Cincinnati
Cincinnati OH 45221-0374
U.S.A.
E-mail: robert.richardson@uc.edu

Abstract. Recent work on self organization promises an explanation of complex order which
is independent of adaptation. Self-organizing systems are complex systems of simple units,
projecting order as a consequence of localized and generally nonlinear interactions between
these units. Stuart Kauffman offers one variation on the theme of self-organization, offering
what he calls a statistical mechanics for complex systems. This paper explores the explana-
tory strategies deployed in this statistical mechanics, initially focusing on the autonomy
of statistical explanation as it applies in evolutionary settings and then turning to Kauffmans
analysis. Two primary morals emerge as a consequence of this examination: first, the view that
adaptation and self-organization should be seen as competing theories or models is misleading
and simplistic; and second, while we need a synthesis treating self-organization and adaptation
as geared toward different problems, at different levels of organization, and deploying different
methods, we do not yet have such a synthesis.

Key words: complexity, development, emergence, evolutionary explanation, explanation,

probability, self-organization, statistical explanation

1. The search for laws of form

Stuart Kauffmans The Origins of Order, promises to reveal the principles

underlying complex and beautiful ordered systems (1993: 173), based on
simple and general principles of self organization (cf. Kauffman 1995). The
work is carefully situated in a tradition that emphasizes laws of form
and their role in explaining biological complexity. The problem biologists
encounter in Kauffmans work concerns how it speaks to the organisms they
encounter and to their evolution. Gabriel Dover observes in a review of
Kauffman that the central issue is at what point do Kauffmans statistical
structures bear upon evolving, historically processed genomes and onto-
genies as we know and love them? (1993: 704) I here offer an answer to
Dovers question, though one that is unlikely to find favor with the friends of
complexity.
656

Discussion of the laws of form has its roots in non-Darwinian

approaches to evolutionary theory (cf. Lenoir 1982; Amundson 1998). It is
often associated with emphasizing the role of development, or developmental
constraints, in the evolutionary process. These developmentalist alternatives
date back at least to Geoffroy Saint-Hilaire, Karl Ernst von Baer, and Richard
Owen in the early 19th century, and include as contemporary representatives
not only Stephen Jay Gould, but biologists such as Pere Alberch, Brian
Goodwin, Stanley Salthe, Steven Stanley, and Gerry Webster, as well as
philosophers such as David Depew, Bruce Weber and William Wimsatt.
Development imposes independent limitations on evolutionary change by
limiting the range and significance of biological variation. Sometimes this
is associated with what are called developmental constraints, though this
expression tends to relegate the influence to a secondary role. More broadly,
the idea is that in accounting for large scale patterns, or for macroevolutionary
change, something other than adaptation by natural selection is required.
Even Darwins most famous advocate, Thomas Henry Huxley, recognized the
significance of development and its potential in elaborating a non-adaptive
evolutionary theory. Huxley believed, as did von Baer and Owen, that there
were a number of distinctive animal types and that similarities in develop-
mental plan resulted in a morphological similarity within types (cf. Lyons
1999). This unity of type led Huxley, at least initially, to doubt the centrality
of natural selection in explaining the patterns of life, and to explore other
explanations for major transitions in the history of life.
The neo-Darwinian paradigm that has oriented most biological research
in the last century recognizes various factors affecting evolution; yet, it sees
only natural selection as playing an important role in shaping the form and
behavior of organisms. Neo-Darwinians hold that, as a reasonable approx-
imation, evolutionary change is a process of adaptation; other evolutionary
factors constitute at most weak constraints on evolutionary change. These
other factors are responsible only for the noise and not the pattern. Accord-
ingly, large scale evolutionary patterns and change the sort of diversification
that determines evolutionary lineages, defining major taxonomic groups are
but small changes extrapolated over enormously long times.
Like many in the developmentalist tradition, Kauffman sees much organic
form as largely independent of adaptation. Though Kauffmans thought has
a decidedly developmentalist cast, it has a unique voice that distinguishes it
from others in the tradition (cf. Auyang 1998). Where neo-Darwinian biolo-
gists see adaptation, Kauffman sees spontaneous or emergent order, arising
from self organization. Where neo-Darwinian biologists see gradual change
driven by natural selection, Kauffman sees phase transitions between neigh-
boring attractors. The method Kauffman deploys is also distinctive. Kauffman
 657

relies on Boolean networks in probing the dynamics of complex systems.

Within these models, stable types emerge as a result of multiple interactions
between large numbers of simple units. Within these models, large scale order
is discernible as the statistically common pattern within sophisticated and
well defined mathematical models.
The statistical character of Kauffmans argument raises a number of issues
which are of independent philosophical and biological importance. More
generally, the evidently stochastic character of evolutionary theory raises a
number of important issues. One concerns the use of statistics in evaluating
evolutionary explanations. As with all uses of evidence, the data relevant to
evolutionary explanations are subject to substantial uncertainty, and infer-
ences concerning the correct explanation are similarly uncertain. Another
equally fundamental issue, and one more directly relevant to Kauffmans
work, concerns the role of probabilistic principles in evolutionary explana-
tions. This raises questions concerning the independence or autonomy of
statistical inference and explanation, as well as the place of probabilistic laws
in explanation. Neither issue is limited to evolutionary theory, though that
is the main context in which I will be discussing them. Kauffmans vision
will occupy me in the two final sections of this paper. Before assessing
the explanatory import of his views, it will be useful to turn to the role
of probabilistic principles in evolutionary explanations. This will put us in
a better position to evaluate the significance of Kauffmans variant on the
developmentalist position.

2. Dualing explanations

Wesley Salmon (1989: 183 ff.) tells the following story. A physicist on an
airplane was sitting near a young boy holding a helium filled balloon, waiting
for takeoff. He asked the boy what would happen to the balloon when the
plane took off, and the boy answered that the balloon would move toward the
rear of the cabin. This seemed right to most of the adults that were there. The
physicist predicted that, to the contrary, it would move forward. It did. Why?
There are two explanations rather than one. On the first, we know that as the
plane accelerates, the rear wall of the plane creates a pressure gradient from
the back to the front of the plane. Because there is more pressure toward the
rear, the bombarding molecules move the balloon forward. On the second,
we know that an accelerating reference frame is equivalent to an inertial
frame within a gravitational field. These states of motion are equivalent.
Helium balloons tend to move away from the source of a gravitational field,
as we all know. So the balloon moves forward, away from the equivalent of
a gravitational field. These are two explanations. Salmon observes that the
658

first appeals to proximate causes involved in pushing the balloon forward,

while the second appeals to very general physical principles without relying
on causal mechanisms. As a way of marking the difference between them, we
may think of the former explanation as causal or causal/mechanical and the
latter as, broadly, structural. Which is the right explanation? Both are correct.
As Salmon says, both explanations are illuminating in their own way. They
are also different.1
Lets try a biological example one that enforces a similar moral. It has
long been noticed that a 1:1 ratio of males to females is common. Indeed,
the ratio of male to female births in humans (at 19:18), carefully adjusted
so that it would maintain a 1:1 ratio among adults and thereby monogamous
marriage, was taken in 1710 by Arbuthnot as a signal of Gods forethought
(cf. Gigerenzer et al. 1989). We may be skeptical about the appeal to Gods
design in explaining a 1:1 ratio, but the ratio is certainly in need of explaining.
R. A. Fishers (1930) explanation of it is now nearly canonical, and is broadly
accepted in the biological community. Fisher argued on general grounds that
the optimal reproductive strategy under a wide range of conditions is to invest
equally in males and females. A 1:1 ratio is a stable optimum, and under
natural assumptions no other ratio is stable.2 Darwin evidently came very
close to this explanation in The Descent of Man:
Let us now take the case of a species producing . . . an excess of one
sex we will say the males these being superfluous and useless, or
nearly useless. Could the sexes be equalized through natural selection?
We may feel sure, from all characters being variable, that certain pairs
would produce a somewhat less excess of males over females than other
pairs. The former, supposing the actual number of the offspring to remain
constant, would necessarily produce more females, and would therefore
be more productive. On the doctrine of chances a greater number of the
offspring of the more productive pairs would survive; and these would
inherit a tendency to procreate fewer males and more females. Thus a
tendency towards the equalization of the sexes would be brought about
(1871: 316).
Darwin did not generalize this argument to show how a population producing
an excess of females would be subject to similar correction. An excess of
females would, after all, be neither superfluous nor useless. I expect this
simple fact explains why Darwin did not extend the point successfully.
Fisher did see how the generalize it. The central feature of his analysis
requires that we look at more than one generation. Offspring generally can be
viewed as making contributions to parental fitness, depending on the number
and fitness of the offspring sired. Reproductive strategies may vary in the
number of offspring, or in the fitness of offspring sired. Both survivorship
 659

and reproductive potential contribute to their fitness values. In some circum-

stances, having more offspring is better; in other circumstances, having fewer
offspring is better, particularly if this increases their likelihood of survival.
Likewise, reproductive strategies may vary in the proportion of offspring of
various types. Fishers problem concerns specifically the proportion of males
and females and how variation in relative investment affects parental fitness.
If on average, a population invests more in females than males, then there
will be a differential advantage to individuals that invest more heavily in
males because the male offspring will be at a reproductive advantage in the
next generation. Investing in males is better than investing in females when
females are more common. By contrast, if on average a population invests
more in males than females, then there will be a differential advantage to
individuals that invest more heavily in females because those females will
be at a reproductive advantage in the next generation. Investing in females
is better than investing in males when males are more common. That is,
individuals investing in a way deviating from the norm will have offspring
that are at a reproductive advantage if the population is not at an equilibrium.
Given that the tendencies are heritable, selection should favor a 1:1 ratio.
Fisher gave this a brief but elegant formal characterization. If males and
females require an equal reproductive investment to bring them to maturity,
the ratio of the two sexes defines the relative fitness of the two reproductive
types. If the cost to produce a son is cm and the cost to produce a daughter is
cf , and if the percentage of sons is p, then the total benefit accruing to a parent
will be a function of the benefits due to sons, bm , and to daughters, bf :

bm pT/cm + bf (1 p)T/cf

An alternative investment in sons of p* will result in a differential advantage

insofar as

bm p T/cm + bf (1 p )T/cf

is greater than the benefit to any alternative strategy. If sons and daughters are
equally costly, then sons will provide a greater return when females are more
common, and daughters will provide a greater return when males are more
common. The equilibrium point is one in which the average numbers of sons
and daughters is the same.
It is worth underscoring that this general result enables us to explain
both the general phenomenon, as well as the individual cases. A 1:1 ratio is
selectively favored under such a broad range of circumstances that we would
expect it to be common. Assuming the broad parameters defining the model
are common, the mathematical result explains why a 1:1 ratio is common.
660

Moreover, if we are asked why there is a 1:1 ratio in humans, that is explained
in exactly the same way. If the explanation is a sound one, it explains not only
why a 1:1 ratio is common, but also why we encounter a 1:1 ratio in particular
cases.
Elliott Sober (1983, 1984) calls such explanations equilibrium explana-
tions, contrasting them explicitly with causal explanations. He says:
Equilibrium explanation shows why the actual cause of an event is, in
a sense, explanatorily irrelevant. It shows that the identity of the actual
cause doesnt matter, as long as it is one of a set of possibilities of a
certain kind (1984: 140).
Whereas identifying the cause of some state of the population depends on
specifying the actual cause, and knowing the actual history, equilibrium
explanations such as Fishers do not. Equilibrium explanations more gener-
ally show that some state is stable: an equilibrium state is insensitive to
modest perturbations; and systems deviating from the equilibrium, at least
within a certain range, will tend to evolve toward the equilibrium state. Equi-
libria are a species of what are called attractors. An equilibrium explanation
is an explanation of why the equilibrium state in this case, a 1:1 sex ratio
is stable and others are not.
Sober says that when we are confronted with equilibrium explanations,
the details of a populations past often do not matter to its present config-
uration (1984: 141). This is right in one sense, but slightly misleading in
another. Equilibrium explanations do not depend on any particular causal
history; that does not imply that there are no causes for what we observe or
that the causes do not matter. Whether we are explaining the generality of a
1:1 sex ratio, or its occurrence in some given species of altricial birds, Sober
is right that Fishers explanation does not cite specific causes and does not
depend on any particular history. As Sober argues, selective explanations are
causal, depending on the extent of variation and the range of factors acting on
the population; Fishers argument is not a causal explanation in anything but
a trivial sense (1983: 203204). Fishers explanation is also general in a way
that selectionist explananations are not. It appeals to no specific cause; indeed,
there likely is no single cause common among the entire range of cases.3
Philip Kitcher adopts a similar position. Kitcher says that, confronted with a
sex ratio very near 1:1, the right explanation of this is that there are selection
pressures that favor the evolutionary equilibrium. Though we might be able to
detail a complete causal history based on sperm and egg production, mating
patterns, and the like, this information is, in a way, irrelevant to what we want
to explain. He says . . . the causal approach seems to err by overlooking the
fact that the particular phenomenon to be explained is one example of a class,
all of whose members instantiate a general regularity (1989: 426). Again
 661

this is right in a sense and may be misleading in another. It is true that the
explanation does not depend upon any more detailed causal history, whatever
that might be; and it is true that the explanatory force depends on the general
regularity which is present; however, any more detailed causal history would
be irrelevant to the explanation of the ratio even when what we are explaining
is the specific case and not the general regularity.
Sober and Kitcher each recognize an important feature of such explana-
tions, one that I want to exploit. Equilibrium explanations abstract from the
causal history. The explanations do not depend upon any specific causal
history. Equally, they do not depend upon local mechanisms. There are, we
know, a variety of mechanisms for determining sex. In humans, sex deter-
mination depends largely on XY chromosomal distributions and how they
segregate in the formation of gametes. In other organisms, sex is determined
quite differently. Fishers explanation abstracts from such details as much as
from causal history. We may be sure that there are such local mechanisms,
just as we may be sure that there was some history relevant to its evolution.
Fishers explanation does not depend on such niceties, and would not benefit
from including them.
Genetic drift provides an analogous case, insofar as it abstracts from
specific causes whether local or historical and emphasizes the pattern.
Natural selection and drift both can explain the dynamics of populations,
how gene frequencies, or genotypic frequencies, or phenotypic frequencies,
change over time. Evolution is clearly a stochastic process in the sense that,
given an initial distribution of genes, or genotypes, or phenotypes, with real-
istic parameter values we are at best able to project a probability distribution
of the relevant states. This is associated with genetic drift. Genetic drift is
simply the error in transmission of types from generation to generation,
arising from finite population size and inbreeding (cf. Wright 1932). Drift
is standardly treated using models incorporating ensembles of finite popula-
tions. Effectively, we ask what happens if we have the same evolutionary
problem repeated. Given a single gene with two alleles at frequencies pi
and pj , then in the absence of selection ensembles of populations initially
polymorphic at that locus will tend to disperse across a wide range, from
populations fixed for one allele to populations fixed for the alternative allele
(see Falconer 1989 or Roughgarden 1979). These changes are random within
each subpopulation if there is no selection operating, and as a result different
subpopulations become differentiated though there will be no change in
the allelic frequencies in the overall population. This differentiation among
subpopulations is a chance affair. Since the extremes in which subpopula-
tions become fixed for one allele or another are absorbing states (at least in
the absence of mutation or immigration), as the ensemble disperses over the
662

Figure 1. Drift in a population without selection. This illustrates the effect of drift in a popula-
tion in the absence of selection. The x axis represents the frequency of the A allele within
populations, and the y axis represents the proportion of populations within the ensemble.
Initially, all the populations in the ensemble have 60% A alleles (and therefore 40% of the
alternative). Over time, the populations disperse, with some reaching fixation for A and
some becoming fixed for the alternative allele. In the absence of mutation or outcrossing,
populations which become fixed for an allele remain fixed. The rate at which populations
become autozygous is inversely proportional to the effective population size. At the limit, all
populations become fixed for A or an alternative. Given that initially the frequency of A is
60%, the expected outcome is that 60% of the populations become fixed for A.
 663

space each population will eventually become monomorphic. In the limit,

drift ensures that the ensemble of populations will bifurcate into a bimodal
distribution at the two extremes (see Figure 1).
Drift can explain both the general phenomenon and the individual case.
Given an initial degree of polymorphism, we have a precise prediction of the
pattern we should expect: the overall frequency of the genes should remain
unchanged, while the populations become increasingly monogenetic. The
frequencies of populations fixed for the alternative alleles, more specifically,
should be the same as the initial frequencies of the alleles. So if we begin
with pi and pj as the frequencies of the alternative alleles, in the absence
of selection and mutation those frequencies should remain unchanged in the
metapopulation, even though the subpopulations become genetically uniform.
Moreover, the frequency of populations fixed for the i and j genes should,
respectively, be pi and pj . Equally, we have an explanation applicable to
individual cases: a particular population has a determinate probability of
becoming fixed for one allele as opposed to another. If a given population
becomes fixed for a given allele, we can explain this as a consequence of the
consequence of drift within a population having some frequency of the allele,
and the rate at which this happens is a consequence of population size. If we
begin with i and j at the frequencies pi and pj , and a population becomes fixed
over time for the j allele, then it perfectly proper to explain this as something
which has a probability pj of happening.
The neutral case is an ideal, assuming that there are no selective differ-
ences. If we focus on selection alone, in the absence of drift, immigration and
mutation, we have a deterministic process: given a frequency distribution at
one time, and fixed selection coefficients, then there should be a unique distri-
bution of frequencies in the next generation. Mathematically, this is modeled
using infinite population sizes in which sampling error, and therefore drift,
could not occur. In the deterministic case, the change in the frequency of a
trait with a value zt at t will be

zt = zt+1 zt h2 2 (szt zt )

where s is the selection coefficient and h2 2 is the heritable variance. In

an infinite population, these changes would be a deterministic function of
fitness. Under directional selection in a population short of equilibrium, the
result is in an increase in the frequency of more fit individuals, and in average
fitness (see Figure 2). With finite populations, drift has the effect of exploring
adaptive zones. In finite populations subject to selection (cf. Lande 1976;
Sewall Wright 1931, 1932), drift captures the extent to which actual changes
in gene, or genotypic, or phenotypic frequencies tend not to be correlated
with fitness differences; the variation around the norm defined by selection is
664

Figure 2. Changes in frequency due to selection. This illustrates the expected change in a trait,
or the frequency of a gene, under directional selection and in the absence of drift. The change
from one generation to the next is a function of the heritable variance, and the differences in
fitness (or the selection differential). If selection is the only factor operating, then fitness at
t + 1 is a deterministic function of the fitness at t.

random with respect to fitness. Fitness values then determine the strength and
location of the central tendency within an ensemble of populations, and drift
becomes the amount of dispersal around the mean value (cf. Richardson and
Burian 1992; this is illustrated in Figure 3). The extent of the variation the
scatter around the norm is a function of population size.
Once again, the explanation of such variation is abstract. As with equi-
librium explanations, the actual causes are explanatorily irrelevant. They
are nonetheless consistent with causal explanations. Fishers explanation
is consistent with any number of historical scenarios and any number of
selective regimes: so long as an evolutionary trajectory results in a 1:1
ratio because a 1:1 ratio was favored, then whatever the specific cause, it
will conform to Fishers model. Explanations in terms of drift are likewise
consistent with any number of historical scenarios. In fact, the range of causes
must be heterogeneous, if the variance around the curve defined by selection
is random. The point is not merely that we do not know the specific causes.
It is not even that we do not need to know the specific causes. Including them
would not be a useful supplement to the explanation we already have. We do
not know what the causal or historical antecedents were. We do not care.
It would be wrong to think that these explanations are somehow incom-
plete, to think of them as surrogates for causal explanations, to insist on
 665

Figure 3. Changes in frequency under selection with drift. The mean expected change is a
function of the heritable variance and the selection differential, as in the deterministic case.
The variance in the frequency changes around that expected value are indicative of the signifi-
cance of drift, and is inversely proportional to the effective population size. The top figure
illustrates the larger population size, in which the effects of drift are diminished. The bottom
figure illustrates a smaller population size, with the effects of drift correspondingly enhanced.
666

recasting them as causal explanations, or to project their reduction to causal

explanations. Fishers model is abstract but not incomplete, and it is not a
surrogate for some ideally complete causal explanation. The point can be
illustrated by asking whether we would retain Fishers explanation as an
explanation of 1:1 sex ratios, were we to detail the causal history in the
evolution of the ratio in a specific case. If we knew that a given population had
significantly more males than females, say, and knew that there was intense
competition for mates among those males, we might come to know what
specific advantages favored some among the many males. This might be an
important evolutionary question, but it would not answer the question with
which we began. Fisher can explain not only why this population, in these
circumstances, has a 1:1 ratio, but why it would have been expected even
had the circumstances been different than they in fact were. To shift the case
slightly, suppose we were to discover that a 1:1 sex ratio is ancestral in a
given population, rather than derived. In that case, no such causal explanation
is available at all. The 1:1 ratio is not derived, at least within the population.
One dogmatically committed to a causal theory of explanation would then
be committed to rejecting the equilibrium explanation if these are supposed
to be incomplete causal explanations or surrogates for causal explanations.4
Fisher would disagree, and there seems to be no reason to reject Fishers
explanation here if we accept it in any case at all. Likewise, an explanation
of some evolutionary change in terms of drift abstracts from any particular
causal history. What it provides is an abstract framework in which to deal with
changes in the frequency of genes, genotypes, or phenotypes. The explanation
is fundamentally and irreducibly stochastic. It begins with the more general
patterns of change, or expected patterns of change, and proceeds to the more
specific cases. Once again, the actual causal history does not matter. We may
know, for example, that a population has undergone some change in the distri-
bution of genes, perhaps bringing one allele to fixation. Given parameters
on the population, we can determine the probability with which this would
happen as a result of chance alone and the rate at which this should happen
(cf. Lande 1976; Falconer 1989). This is a perfectly acceptable explanation
from an evolutionary perspective, even though it does not invoke the specific
causes in a given case. Again, this does not deny that there are causes; the
initial point is that the details of the causal history are not necessary to explain
the result. The adequacy of the explanation depends entirely on whether it
adequately captures patterns of change, and how a specific instance fits within
these more abstract patterns. This is not an equilibrium explanation, but like
equilibrium explanations, the significance of drift depends on its ability to
explain changes because they conform to a more general pattern. Again, no
 667

causal history features in the explanation. None is necessary. Again, we do

not know what the causal or historical antecedents were. We do not care.
We can see this explanatory strategy clearly at work in actual cases. In
the classic studies of blood group distributions in Italian populations in the
Parma valley by Cavelli Sforza and his collaborators (1969), we find the
pattern predicted by drift, given the assumption that blood types are (more
or less) neutral. Within the Parma valley, there is a considerable variation in
patterns of habitation, from smaller less mobile populations in the mountains
to larger more mobile populations on the plain. As one would expect, the
smaller villages are more divergent in blood group types than are the villages
in lower elevations. Though the population of the valley as a whole is very
nearly in Hardy-Weinberg equilibrium, the variation in blood group types
between villages is much higher in the mountains than in the hills, and virtu-
ally disappears on the plain. This is the pattern we should expect with neutral
variation and drift. From the population values, we also can explain why a
specific village has, say, an excess of blood type A. Given the frequencies
of A in the population as a whole, we know the probability that a given
village will become fixed for blood type A. The frequency of the genotypes
in the metapopulation determines the probability of fixation, in the absence
of immigration or selection. The rate at which this happens depends on the
size of the village. Given the overall distribution, we treat individual, isolated,
villages as samples drawn from the population.5
There is, of course, another explanation for the individual cases. Drift is a
consequence of inbreeding. In this case, we can look at the social structures
within the village, ultimately constructing a pedigree which will also explain
why the population became fixed for a specific allele. Indeed, Cavalli-Sforza
and his collaborators did nearly that, studying the parish registers and gauging
the frequencies of consanguineous marriages. With a pedigree in hand, we
would have another explanation of blood type frequencies in particular cases.
(Given such pedigrees for all the villages, we have another explanation for
the pattern as a sum of the individual cases.) This explanation is not incom-
patible with an explanation in terms of drift. One is pitched at the level of the
population and the other is aimed at the more specific question of the result
within a village. One appeals to drift, and the other to the mechanism lying
behind it. As in the case with which we began, both explanations are correct.
They are also different.

3. Kauffmans project

We are now in a position to return to Kauffmans project. One of the central

themes running through The Origins of Order is that the order we see in
668

organisms largely reflects spontaneous order in complex systems (1993: 173).

The idea is the elegant one that systems which are composed of an array of
simple elements, with modest numbers of connections between them, will
spontaneously tend to assume a regularly ordered form. There are striking
examples of this to be found in animal aggregations (cf. Parrish and Edelstein-
Keshet 1999). Schooling fish form patterns with density profiles and move-
ment that persist even with turnover in the group, depending primarily on
simple mimetic behavior. These schools behave as integrated units, some-
times exhibiting arresting visual patterns. The order in these systems is often
described as emergent.6 Similar behavior is exhibited by a wide variety of
what are otherwise apparently very different systems, including genetic regu-
latory networks, protein metabolism, neural networks, ecological systems,
and economic systems. All are complex systems, whose order and behavior
is rooted, Kauffman tells us, in self organization.
There is some ambiguity concerning how ambitious the program of The
Origins of Order is. In some places, the mood is conciliatory:
Unlike physics, biology is the domain of natural selection. If we begin to
glimpse origins of order in the inherent properties of complex systems,
then we must come to understand the mutual interactions of such self-
ordered properties and the actions of selection. Even the outline of an
adequate theory is not available but must be found, for the proper marriage
of these conceptual realms constitutes the proper evolution of our theory
of life (1993: 644).
Natural selection and self organization become complementary explanations.
In the end, we do not think of selection and mutation as competing forms of
explanation, but as factors affecting evolutionary change. The corresponding
task is to understand how self organization and selection can be integrated
into a single scheme. In a similar mood, Kauffman writes this:
The task of enlarging evolutionary theory would be far from complete
even if we could show that fundamental aspects of evolution and ontogeny
had origins in some measure reflecting self-organizing properties of the
underlying systems. The present [neoDarwinian] paradigm is correct in
its emphasis on the richness of historical accident, the fact of drift, the
many roles of selection, and the uses of design principles in attempts to
characterize the possible goals of selection. Rather, the task must be to
include self-organizing properties in a broadened framework, asking what
the effects of selection and drift will be when operating on systems which
have their own rich and robust self-ordered properties (1993: 2223).
In this mood the problem is one of understanding the interaction of selection
and self-organization, as alternative and important sources of natural order.
 669

Natural selection and self-organization complementary rather than competing

principles. In other places, the tone is more ambitious and much less concili-
atory. Thus, in nearly the same breath as he offers consiliatory proposals,
Kauffman suggests:
. . . much of the order seen in organisms is precisely the spontaneous
order in the systems of which we are composed. Such order has beauty
and elegance, casting an image of permanence and underlying law over
biology. Evolution is not just chance caught on the wing. It is not just a
tinkering of the ad hoc, of bricolage, of contraption. It is emergent order
honored and honed by selection (1993: 644).
In this more ambitious vision, self-organization becomes the origin of order,
and natural selection serves the role of a handmaiden, sifting the order that is
there for the asking. This is the more ambitious agenda within The Origins of
Order, one which weds it to the more radical appeal to the laws of form in
developmental biology.
Bruce Weber and David Depew (1996: 44 ff.) give us a range of
possibilities in their useful and insightful discussion of Kauffman:
1. Natural selection, and not self-organization, drives evolution.
2. Self-organization constrains natural selection.
3. Self-organization is the null hypothesis against which evolutionary
change is to be measured.
4. Self-organization is an auxiliary to natural selection in causing evolu-
tionary change.
5. Self-organization drives evolution, but is constrained by natural selection.
6. Natural selection is itself a form of self-organization.
7. Natural selection and self-organization are two aspects of a single
evolutionary process.
These offer a continuum of cases. The first five are conciliatory, acknow-
ledging that selection and self organization are both contributory causes
but differing on their relative importance. The last two offer more radical
alternatives.7
There are different problems affecting the program, depending on how
radical the proposal is. There are problems, moreover, in seeing exactly how
the two are to be combined. In the following section, I want to suggest that,
properly understood, Kauffmans statistical mechanics of complex systems
is aimed at different questions, different explanatory problems, than much of
classical evolutionary biology. Kauffmans problems concern is what is statis-
tically normal, or expected, in such systems. Evolutionary biology is more
typically concerned with the evolutionary history and dynamics of particular
systems. So understood, there is no incompatibility; yet, the moral is not one
of the moderate ones outlined by Depew and Weber. The problems lie in
670

different domains. The methodology makes it unsuited as a competitor to

natural selection. If this is right, then it would be a mistake to somehow view
self-organization as complementary in the way the first five of Weber and
Depews options suggest. It would be akin to thinking of the probabilistic
arrays of statistical mechanics as if they were additional parameters charac-
terizing individual molecules. This leaves us with the more radical options,
about which I will say something in the closing section. The critical sugges-
tion, as Weber and Depew suggest, is that the importance of natural selection
lies in locking in the products of self-organization (1996: 56; cf. Weber
1998). Within the framework of Kauffmans theory, two broad morals are
defended, that it is at the edge of chaos that evolvability is maximized, and
that natural selection achieves and sustains such a poised state. There are, I
will urge, reasons to be skeptical about both suggestions.

4. Self-organization and The Origins of Order

Kauffman offers a formal framework which allows him to pose problems

about the constraints that self-organization imposes on the evolution of
complex systems, and the relationship between self-organization and selec-
tion. He says that what we need is a new kind of statistical mechanics, one
which analyzes the properties of complex systems with very many coupled
elements. It is . . . by understanding the characteristic structure and beha-
viors of the members of such ensembles, that we may be able to understand
both the emergence of order in organisms and its adaptive evolution (1993:
182). Kauffmans approach is partly a matter of necessity. In using computer
simulations, or in theoretical modelling, it is important to focus on the right
level of description. At too small a scale, we are often unable to capture
the dynamics we are interested in understanding. Small scale structure may
simply fail to capture large scale patterns (cf. Goldenfeld and Kadanoff 1999).
We have already seen this general moral at work in the understanding of
population dynamics. It is essentially the insight that led Galton to ignore
petty influences and focus on the statististical pattern. It is the insight that
gives power to Fishers explanation of sex ratios, and the shift that makes
the significance of drift clear. The elaboration of a statistical mechanics for
complex systems is an attempt to shift to a level of description appropriate
for understanding the dynamics of complex systems.
Kauffman essentially brings Boltzmanns insight to the analysis of
complex systems. We predict the systemic properties in terms of the statis-
tically common outcomes of complex systems. This is the key to under-
standing his statistical mechanics for self-organizing systems. The statis-
tical mechanics of the 19th century focused on systems which were inherently
 671

disorganized. In the case of (perfect) gases, we are confronted with systems

of particles, each with six variables describing position and momentum. If
we fix each of these six variables for each particle, this microstate defines a
single point in the phase space for the gas. Since the system is disorganized
that is, the state of the gas is a simple aggregate of its component particles a
sample at one point in the phase space simply wanders over the entire phase
space. Statistically, we can provide a probability that the system will occupy
some volume of the phase space which is described by the macroscopic data.
We need not know the specific distribution of molecules; as we have seen,
knowing that would be no real help in the absence of knowledge concerning
aggregate behavior. The macroscopic phenomena are described as statistical
averages of unknown microscopic aggregates.
In more organized systems, the situation is different. Biological systems
are thermodynamically open systems; moreover, their internal organization
insures that not every region of the phase space is equally open to exploration.
One result is immediate: however intractable the dynamics of classical disor-
ganized systems might be, the intractability increases with highly ordered,
complex, systems. Unlike the simplest cases, in which dependencies are
readily exhibited, in non-linear cases causal dependencies are not obvious
and system behavior can be difficult or impossible to predict (cf. Nicolis and
Prigogine 1993; Wagner 1999). Kauffman attacks the problems these more
highly organized systems present by turning to Boolean networks, systems
of binary variables coupled according to switching functions of arbitrary
complexity. Though such systems are different from disorganized systems,
in many ways there is an analogous moral to be had. Self organization allows
us to describe the features such systems will manifest, and their natural
dynamics. This is done without predicting, or knowing, or caring about,
the specific organization of the system. We abstract from the organization,
and then describe the macroscopic phenomena statistically, as the expected
features of complex systems.
The question and the strategy are important because they display the most
important sense in which what Kauffman provides is a statistical mechanics
of organized systems. Boolean networks offer large numbers of coupled
elements which Kauffman says provide analytically tractable and reasonable
approximations to real systems.8 They are governed by three fundamental
parameters: the number of nodes (N), the number of connections per node
(K), and the number of Boolean functions (P) governing the nodes. The state
of any Boolean node is either 1 (on) or 0 (off). The changes in the system
depend on the states of the input nodes and the Boolean switching function.
That is, the state of a given node at step t + 1 will be determined by the states
of the various input nodes at t and the function that node computes. An and
672

unit will be on in the next cycle provided all its input units are on in the
current cycle. An or unit will be on in the next cycle provided at least one
of its input units is on in the current cycle. The sorts of systems Kauffman
is ultimately interested in are complex, with hundred or thousands of units,
and the Boolean functions depend on multiple inputs. Kauffman sets about
to find what he calls the typical behavior of Boolean networks given these
parameters.
Kauffman simulates networks with varying values for K, varying from K =
N (so that every element is connected to every other element) to K = 0 (so that
none regulates any other). With N nodes, there are 2N possible configurations
of nodes; that is, there are 2N distinct states in the phase space for an N-
node system. A given unit will be on or off in the next cycle depending on
the states of the nodes affecting it and the Boolean function it manifests.
As the number of connections varies, the number of functions increases
dramatically. For K inputs there will be 22K possible Boolean functions. For
each K and N, Kauffman assigns both connections and Boolean functions at
random across the N units. He explains, we generate networks with random
wiring diagrams and random logic, and ask whether orderly behavior emerges
nonetheless (1993: 192). With N nodes and K connections, each node is
randomly connected to K other nodes, which regulate it; and the function that
node computes is randomly assigned from the P possible functions that have
K values. The state of each node at a given time is thus a function of the states
of the K nodes which connect to it; and the state of the system at t (defined
as the configuration of the N nodes, and their activation) is a determinate
function of its state at t 1. The question Kauffman poses is whether there
are any patterns that emerge as typical of such NK systems independently
of the specific patterns of connections we impose.
It is critical to understand that Kauffmans strategy is one that samples
statistically from the range of possible networks with given parameter values
for N, K, and P. He does not attempt to explain the behavior of any of these
randomly generated networks, and there would be little interest in doing so.
None are crafted, for example, so that they have an organization comparable
to any known systems. He asks, instead, what the expected behavior of such
complex systems would be, apart from any preferred order. The expected
behavior is the typical behavior, across the range of possible networks. This
is again a statistical question. The prospect for an explanation based on the
wiring diagrams of these Boolean networks is not a serious one, any more
than we might explain human behavior from a wiring diagram of the brain
if we had one. This is by itself no objection to Kauffmans approach. It is
important to know whether the patterns he has discerned apply equally well
to real systems; and to know this we would need to ask the question whether
 673
Control Parameters:
N: the number of nodes
K: the number of connections per node
P: the number of Boolean functions
Varying values for K, 0 < K < N
2N possible configurations
2 K
2 possible Boolean functions

K=N K values K=1

N
2 mean cycle length N
N
2
e number of attractors N
uncorrelated landscape correlated
disordered character ordered
chaotic dynamics stable

Figure 4. Complex systems and their dynamics. With Boolean functions, NK random
networks are governed by three control parameters: the number of nodes, the number of
connections per node, and the number of Boolean functions. Kauffman systematically varies
K from 0 to N, and finds there are systematic dynamic differences. The characteristics of
networks with K = N and K = 1 are as indicated. The critical values lie at around K = 2, where
systems are neither chaotic nor frozen, and intermediate in the amount of order.

the parameter values he uses are realistic. If they are, then they might charac-
terize the expected behavior for biological systems. Notice that we would not
attempt to describe in detail the behavior of any of these randomly constructed
Boolean networks, and our explanation does not at all depend on being able
to do so.9 Kauffmans results describe the statistically likely result among
systems with these three characteristic control variables.
The generic properties of random NK systems thus do provide a kind of
statistical mechanics for complex systems (Figure 4 provides an overview).
They are properties which would be expected independently of selection,
as a consequence of abstract organization alone. If Kauffman is right, they
provide robust properties across a variety of systems from immune systems
and genetic regulatory circuits to neural networks and economies that are
statistically typical, but not universal. In any given case, the task of detailing
the mechanisms might elude us, but the overall pattern gets an explanation
even if particular cases do not.
If this is right, then the sorts of explanations we get in exploring the space
of NK systems are not incompatible with appeals to selection at all. Neither
are they complementary to selection. Lets consider what Kauffman has to
say about the Cambrian explosion. Roughly 600 million years ago, there
674

was a marked expansion in the range or organisms. The remarkable fauna of

the Burgess shale have left us with a picture of an incredible diversification
of form. Of the thirty five or so living phyla, most or all had their origins
in the early Cambrian. In the Burgess shale, for example, we see sponges,
cnidarians, mollusks, arthropods, echinoderms, hemichordates and chordates,
along with a variety of other phyla some of which are simply bizaar.
Whatever the difficulties there are in interpreting the record (we have, for
example, underrated the precambrian fauna), there is a pattern to be seen that
G.G. Simpson described long ago. Early in the origin of a group, there is a
rapid diversification which is later winnowed down. The surviving lineages
then come to dominate the group. As Simpson saw, the explanation is easy.
Early on, organisms will not be especially well adapted to their environ-
ment, and competition will not be severe. Diversity of form will increase and
become abundant. As they become better adapted (or, in fact, from simply the
vagaries of chance and drift), the number of surviving lineages will decrease.
In the early Cambrian, over roughly 100 million years, we see the emergence
of a wide variety of echinoderm body plans, including eighteen classes; by
contrast, none have emerged in the next 430 million years. Kauffman offers
the following as his take on the Cambrian explosion:
. . . early, poorly fit multicellular organisms could rapidly explore a
large diversity of improved alternative basic morphologies, thereby estab-
lishing phyla. As the rate of finding fitter mutants altering early ontogeny
decreased but fitter variants affecting later ontogeny were still readily
found, variant species founding classes, orders, and the lower taxa became
established. Taxa filled from the top down (1993: 77).
Beginning with a relatively rugged fitness landscape, but in which organisms
are poorly adapted, there will be many variants which increase fitness. Even
those variants which involve modifications in fundamental body types may be
reasonably well adapted. The result would be a rapid proliferation of phyla.
The rate at which we find dramatic changes are fitter, though, decreases very
rapidly as the average fitness increases. So we expect a rapid diversification
followed by more gradual and less dramatic adaptation as organisms become
more fit. I dont wish to quarrel with Kauffmans interpretation (and in any
case, many of the details are up for debate), but want to make a simple
point. Kauffman may be able to explain the general pattern of radiation and
consolidation.10 What it will not explain is the details we see in the Cambrian.
The Cambrian fauna was, for example, limited in the ways they feed, with a
number of suspension feeders, and a few carnivores and herbivores. One of
the most remarkable changes is the reorganization necessary for the evolu-
tion of the arthropod brain. These kinds of details the specific changes
we see in the Cambrian are simply not part of the story Kauffman tells.
 675

He may explain the pattern of radiation and consolidation. But the arthropod
brain is not something he even attempts to explain, much less how they are
related to, say, flatworms. These sorts of details are not within the scope of
Kauffmans vision (cf. Burian and Richardson 1996: 165 ff.). This is less a
criticism of Kauffman, than an observation that his explanatory tools are in
many ways different from those of the classical evolutionary biologist. The
evolution of the arthropod brain, or of skeletalized structures, or burrowing, is
the substance on which classical evolutionary biologists obsess. It is simply
not part of the story of The Origins of Order. Does this mean that Kauffman
is wrong? No. Does this mean that self-organization renders classical evolu-
tionary, adaptive, theory superfluous or redundant? No. Both could in fact be
correct. They are no more in competition than are Salmons two explanations
of the balloons motion.
This provides a partial answer to Dovers question, with which we began.
Understood this way, Kauffmans statistical structures do not bear directly on
the evolution of organic structures. They will not tell us anything of interest
about the evolution of the arthopod brain, or explain the particular results we
see following the Cambrian explosion. The most they tell us is something
about the more general patterns we might expect to observe in, say, adaptive
radiation or the sorts of organization we might expect to find in genetic
regulatory systems. Kauffmans statistical structures might tell us something
about the statistical patterns; for a grasp of the particular results, we need a
more detailed grasp of the systems at hand. I emphasize that they might shed
light on these more general patterns. Whether they actually do so will depend,
in part, on whether the models depend on realistic parameter values.

5. Adaptation to the edge of chaos

This leaves us with Kauffmans more radical proposals. What can we say
concerning the proposed marriage of self organization and selection? The
Origins of Order, after all, promises to reveal the principles underlying
complex and beautiful ordered systems (1993: 173), based on simple and
general principles of self organization. So far, Ive pressed that The Origins
of Order and The Origin of Species are not in competition. As Ive recounted
it, Kauffmans central strategy relies on the idealization of complex Boolean
networks with multiple couplings. These systems are governed by at least
three parameters: the number of nodes (N), the number of connections per
node (K), and the number of Boolean functions governing the nodes (P). He
considers networks with varying values for K, from K = 0 to K = N. For
a given N and K values, Kauffman assigns both connections and Boolean
functions at random across the N units. Effectively, he samples from among
676

the space of possible random NK systems, attempting to discern the patterns.

He asks whether, even with such random networks, some orderly behavior
emerges. If there is insight to be gained concerning complex systems from
such a statistical mechanics, it is not at a level of detail competing with clas-
sical evolutionary explanations. It must be pitched at a more abstract level.
We must proceed without attending to the petty influences that affect the
behavior of individual systems. What does Kauffman have to offer us here?
For the sake of concreteness, lets treat the N nodes as genes with alterna-
tive alleles, and consider how to treat fitness as a function of the number
of loci and their interactions.11 With K = 0, there is exactly one optimum
genotype, composed of the aggregate of the N more fit alleles. All other geno-
types will be less fit. Moreover, for any genotype with m less fit alleles, any
alteration among those m alleles will bring us closer to the global optimum.
Any suboptimal genotype can climb to the fitness optimum by m steps, substi-
tuting the alternative alleles for the m less fit alles. Each substitution yields
an incremental increase in fitness. This is a smooth, highly correlated, fitness
landscape. It is the perfectly additive case, in the sense that interaction effects
do not exist. If N is small, then the increment in fitness for any particular
change may be large, but it is independent of the other nodes. The landscape,
though correlated, then will be steep. If N is large, then on average the incre-
ment in fitness will be small. Each gene will make some difference, but most
will be minimal. The landscape will be gradual. In the latter case, since fitness
differentials between alternatives are relatively small (at least on average),
they might be balanced by mutation pressures. In the former case, relatively
small changes will result in large displacements.
Consider the other extreme, where K approaches N. These are integrated
networks, in which there is interaction between all the elements. In the limit,
when K = N, every node is connected to every other. Any change in any node
affects everything else. The contrast with the K = 0 case is marked. Whether
or not there is one global optimum, there will be a variety of local optima.
It will, moreover, be difficult to find the global optimum by incremental,
gradual, change; and there will be a tendency to be trapped on local optima.
A gradual adaptive walk will miss the global optimum, getting trapped on
lower peaks. Kauffman shows that the expected number of local optima is, in
fact 2N /(N + 1). As N increases, the number of local optima increases expo-
nentially. In the most extreme cases, fitness values are uncorrelated: a mutant
differing at even one locus from its parent will differ randomly in fitness
from that parent. There are two dynamic consequences Kauffman highlights.
First, adaptive walks will tend to reach only a few local optima. Any generic
order should be visible. Second, since there is massive interaction, there will
be conflicting constraints on every gene. Given these conflicting constraints,
 677

adaptive peaks will fall toward the mean fitness of the ensemble. Even the
best will not be very good. The optimal genotypes become hardly better than
chance genotypes in the space of possibilities (Kauffman 1993: 5253).
Kauffman says that this sort of tradeoff reveals a fundamental restraint
facing adaptive evolution (1993: 53). The dynamics depend on the N and
K values. If K is nearly the same as N, then there will be a number of local
optima, but any one genotype will only be marginally better than the average
and will not resemble its neighbors much more than more distant relatives:
the landscape, Kauffman says, is uncorrelated. If K is much smaller than
N, then there will be high optima, but any genotype will be only marginally
better than its neighbors in the space of genotypes. The differences between
neighbors will be much less than the average differences: the landscape is
highly correlated.
In the most general terms, what Kauffman found is that with a K value near
N, the system is maximally disordered and chaotic that is, highly sensitive
to initial conditions and perturbations. As one would expect on probabilistic
N
grounds alone, cycles lengths are on average ( 2 )/2, where there are 2N
possible states. (This is simply the number of steps at which the probability of
finding one of the possible states by random search is 0.5.) With large values
of N, this is an enormous number. Since the average cycle length is rather
long, the number of state cycles which will form is small. Finally, change
is chaotic in the sense that the state of the system changes unpredictably
with small changes in components. Even with values of K much below N,
Kauffman tells us that these general features are retained. As K decreases,
though, the spontaneous order is more marked. When K = 2, so that a unit
receives input from only two other units, Kauffman tells us thatthere is a
large amount of spontaneous order. The average cycle length is N , which
means that a Boolean network with 10,000 elements would, on average, limit
itself to 100 states; accordingly, the number of attractors is also N. There
will be spontaneous order. The explanation for this kind of spontaneous order,
Kauffman tells us, is simple. Some Boolean functions are canalyzing func-
tions: they are such that a particular input from a unit necessitates some state
of the regulated unit, independently of the state of other input units. An and
unit is a canalyzing function in this sense, since if one input unit is off, the
regulated unit will be off; similarly, (inclusive) or units are canalyzing, since
if one input unit is on the regulated unit will be on. This has the effect, with
reasonably short state cycles, of freezing some units on and some off. The
result is that the only units changing are functionally isolated from others.
Attractors are small. Change is localized, and as a result the state of the
system changes minimally with changes in components.12 Kauffman says
Random NK Boolean networks with K = 2 inputs to each of 100,000 binary
678

elements yield systems which typically localize behavior to attractors with

about 317 states among the 2100,000 possible alternative states of activity . . .
note and remember that our intuitions about the requirements for order in
very complex systems have been wrong (1993: 235).
Thus, with a K value near N or in fact with K 5 the system is
disordered and chaotic. Moreover, though there are relatively few attractors,
they are highly sensitive to perturbations. At the other extreme, with K =
1, there is tremendous stability and order. The number of attractors is large,
and the effects of perturbations are limited and localized. It is at K = 2 that
Kauffman finds interesting dynamics. There is a large amount of spontaneous
order, with a moderate number of alternative attractors. Here at the edge of
chaos Kauffman sees a serious hope to account for much of the order we
see in natural systems (1993: 202). In the highly ordered regime, systems are
frozen in place and in the chaotic regime, the dynamics are unpredictable.
It is at the boundary that evolvability is at its greatest. Kauffman goes on to
suggest that natural selection achieves and maintains networks poised here
at the edge of chaos. Though he does not consider the support for either
suggestion decisive, he is led to the bold hypothesis that here at the edge of
chaos evolvability is maximized, and natural selection achieves and sustains
such a poised state (1993: 232). David Depew and Bruce Weber describe
the prospect, with apparent approval, saying Adaptability, on this view, is
itself a paradigmatic adaptation. So construed, adaptability depends on main-
taining a great deal of variation in populations. . . . [Natural] selection can be
construed as favoring self-organized genetic arrays as a source of variation
(1995: 484).
Consider first the proposition that it is at the boundary between order and
chaos at which evolvability is maximized. The highly ordered regime is one
in which perturbations generate minimal overall change. In other words, there
is minimal variation. The chaotic regime is one in which perturbations have
unpredictable effects: there is no correlation between the initial and perturbed
states. There is no heritability. The edge of chaos is simply the region in
which there is heritable variation. It is true that heritable variation is necessary
for evolution (cf. Lewontin 1970). It is then true that systems at the edge of
chaos would be evolvable, because it is only here that there is heritable
variation. Kauffmans first hypothesis that evolvability is maximized at the
edge of chaos is true, once seen in this light, though perhaps not as bold as
it was in 1859.
Consider now the proposition that natural selection achieves and sustains
this poised state at the boundary between order and chaos. This now becomes
the proposition that natural selection maximizes heritable variation. The
picture of adaptation as a search process is retained, but now understood as
 679

a process which maximizes evolvability, or adaptability, understood as

the amount of heritable variation. It has long been known that in spatially or
temporally variable environments, there is an advantage to variation. This
might be supposed to support Kauffmans second conjecture. It is, unfor-
tunately, phenotypic plasticity which is generally considered to be adaptive
(cf. Levins 1963 and 1968; Lewontin 1957; Schmalhausen 1949; Stearns
1989; and for experimental confirmation Sultan and Bazzaz 1993a, 1993b,
and 1993c; Day et al. 1994), and its role in modulating the results of selection
changes the picture substantially. Among other effects, phenotypic plasticity
allows adaptation to local environments that reduces the fitness differences
between genotypes, allowing them to converge on adaptive responses (Wright
1931). As a result, though it allows for variability in phenotypic response, it
can diminish rather than increase the heritable variance. We must look else-
where, and specifically to genetic variation, to support Kauffmans conjec-
ture. There are a number of factors affecting genetic variation in populations.
Some factors affecting genetic variation also appear to be subject to selection.
Kauffman says this, in his reconstituted role for natural selection:
Evolvability, the capacity to search a reasonable fraction of the space, may
be optimized when landscape structure, mutation rate, and population size
are adjusted so that populations just begin to melt from local regions of
the space (1993: 95).
Mutation rate is one the the basic parameters in population genetic models
of evolution, as is population size. There are a number of ways to measure
mutation rate. There is considerable variation in mutation rate. Mutation
rates differ by several orders of magnitude between different species, and
can differ dramatically even within the same organism. Among vertebrates,
for example, mutation rates of mitochondrial DNA are three times as great
as the rate for nuclear genes. By contrast, in plants, the mutation rates for
nuclear DNA are five to ten times as great as the rate for mitochondrial
DNA (Mitton 1994). It is not surprising to find that in animals mitochondrial
DNA evolves faster than nuclear DNA, and conversely in plants. Moreover,
there are a number of DNA repair mechanisms, which serve as evidence of
natural mechanisms to control the effects of mutation. It does appear to be the
case that mutation rates (and repair mechanisms) are under selective control.
Kauffmans conclusion is that mutation rates are modulated to maximize
heritable variation. Theoretical models, however, indicate that this is not
generally true (Gillespie 1981a and 1981b; Holsinger and Feldman 1983;
Holsinger et al. 1986; Lieberman and Feldman 1986). If we assume that
there is a modifier gene controlling the rate of mutation in target genes,
with different alleles yielding different rates of mutation, then the different
alleles will result in different variances in fitness. The result is simple. With a
680

constant environment, and a constant assignment of fitnesses, selection favors

a lower mutation rate. In the models, alleles with lower mutation rates always
go to fixation. Under these conditions, at least, selection minimizes variation.
The result is not hospitable to Kauffmans conjecture. Kauffmans second
bold conjecture that selection favors systems poised at the edge of chaos
and thus maximizes evolvability is indeed a bold one, but it is one we should
greet with polite skepticism.

Acknowledgements

This paper has emerged over a number of years, and with the encouragement
of a large number of people. That does not mean, of course, that they agree
with the conclusions. Among them, I would especially like to thank William
Bechtel, Robert Brandon, Richard Burian, Lindley Darden, David Depew,
Peggy Des Autels, Peter French, Miriam Solomon, and Stuart Kauffman.
This work was supported financially by the Taft Fund at the University of
Cincinnati.

Notes
1 It is, more generally, a central project of classical physics, whether Newtonian or post-
Newtonian, to describe and explain the trajectories of physical particles, whether these are
under the influence of forces (for example, gravitational or electromagnetic) or free from such
influence. Within Newtonian physics, gravitational forces are typically treated within a three
dimensional space with an absolute time. The gravitational force acting on a particle is then a
three dimensional vector field. Michael Friedman (1983) shows that it is possible to dispense
with such forces within a four dimensional space-time. Falling particles follow the geodesics
within this reformed space-time. These also are not incompatible explanations, though both
are structural in some sense.
2 Those assumptions are not always met, and deviations from the 1:1 ratio are explained
accordingly. The central case is the skewed ratio of females to males in the social insects,
but there are dramatically divergent ratios in other species as well. These can generally be
explained by the way they defy the assumptions which drive Fishers argument. It is, in fact,
one of the attractive features of Fishers argument that it explains the prevalence of 1:1 ratios
while also making the deviations intelligible.
3 Explanations which depend on showing something is an evolutionarily stable strategy
provide clear cases of optimality arguments, which are a species of equilibrium explanations
(see Maynard Smith 1975, 1976, 1982). These have been applied broadly, including sexual
behavior and animal aggression. The emphasis on stable optima has the consequence that
such explanations do not offer us any immediate insight into evolutionary history; that in turn
provides some reason to be skeptical about them as selective explanations (see Richardson
1984).
4 A causal theorist has substantial lattitude here. The prevalence of a 1:1 ratio, or its presence
 681

in a given species, might be explained causally as the result of stabilizing selection, even if a
1:1 ratio is ancestral. This is independent of the issue I am raising, which would survive even
if there is no variation, and hence no selection, for a 1:1 ratio.
5 Cavalli-Sforzas studies are often cited as if they confirm the significance of drift. As
an historical matter, this may be misleading. Sewall Wright had earlier claimed that the
major blood group differences had no adaptive significance. Subsequent empirical studies
showed that there were in fact different susceptibilities to diseases, with some afflictions more
commonly associated with one or another blood group. In this context, Cavalli-Sforzas work
demonstrates not the importance of drift, but the neutrality of blood groups assuming the
significance of drift (for a fuller discussion, see Gigerenzer et al. 1989: ch. 4).
6 For discussion of the concept of emergence as it applies here, see Bechtel and Richardson
(1992), and Bechtel and Richardson (1993), ch. 9.
7 In Darwinism Evolving (1995). Depew and Weber offer a similar range of possibilities, and
express sympathy with Kauffmans approach, emphasizing the more radical view that natural
selection maintains systems at the edge of chaos. See Depew and Weber (1995), chapter 16,
especially pp. 454 ff. Richard Burian and I (1992) recognized the more radical and more
conservative proposals, and defended the importance of the third option above.
8 This may turn out not to be an innocent idealization. Boolean networks are discontinuous,
and with continuous activation functions some attractors turn out to be unstable. Also see
Richardson (1997) for a discussion of this pertaining to genetic regulatory systems.
9 This does not mean that the behavior of such systems is necessarily well understood. For
example, it would be possible, with higher values of K, for spontaneous order to emerge, with
a restriction on the number of Boolean functions which are used. Again we would end up with
forcing structures and frozen components. Determining what count as reasonable parameter
values is an important project. This does not carry us, though, to the project of understanding
the detailed organization of individual systems. We still remain at the more abstract level,
focused on the statistical behavior and how it varies with changes in parameter values.
10 How much follows from his general principles and how much is just redescription in his
own terms is something I leave open.
11 Kauffman follows his characteristic strategy here, selecting one configuration one point
in the phase space arbitrarily as defining the most fit.
12 T. W. Zawidzki (1998) explores the differences between Kauffmans models and Herbert
Simons reliance on (nearly) decomposable systems. Though Zawidzki is right to emphasize
the differences between Simon and Kauffman in terms of the range of architectures, K values
at or near 2 are nearly decomposable.

References

Amundson, R.: 1998, Typology Reconsidered: Two Doctrines on the History of Evolutionary
Biology, Biology and Philosophy 13, 153177.
Auyang, S.Y.: 1998, Foundations of Complex-system Theories, Cambridge University Press,
Cambridge.
Bechtel, W. and Richardson, R.C.: 1992, Emergent Phenomena and Complex Systems, in
A. Beckermann, H. Flohr and J. Kim (eds), Emergence or Reduction? Walter de Gruyter,
Berlin, pp. 257288.
Bechtel, W. and Richardson, R.C.: 1993, Discovering Complexity, Princeton University Press,
Princeton.
682

Burian, R. and Richardson, R.C.: 1991, Form and Order in Evolutionary Biology, in A. Fine,
M. Forbes and L. Wessels (eds), PSA 1990, Philosophy of Science Association, Volume 2,
pp. 267287.
Cavelli-Sforza, L.L.: 1969, Genetic Drift in an Italian Population, Scientific American 223,
2633.
Darwin, C.: 1871, The Descent of Man, and Selection in Relation to Sex. Reprinted with an
introduction by J. T. Bonner and R. M. May: 1981, Princeton University Press, Princeton.
Day, T., Pritchard, J. and Schluter, D.: 1994, A Comparison of Two Sticklebacks, Evolution
48, 17231734.
Depew, D.J. and Weber: B.H.: 1995, Darwinism Evolving, Bradford Books/M.I.T. Press,
Cambridge.
Dover, G.: 1993, On the Edge, Nature 365, 704706.
Falconer, D.S.: 1989, Introduction to Quantitative Genetics, 3rd edition, Longman, Essex.
Fisher, R.A.: 1930, The Genetical Theory of Natural Selection, Oxford University Press,
Oxford.
Friedman, M.: 1983, Foundations of Space-Time Themes, Princeton University Press, Prin-
ceton.
Gigerenzer, G. et al.: 1989, The Empire of Chance, Cambridge University Press, Cambridge.
Gillespie, J.H.: 1981a, Mutation Modification in a Random Environment, Evolution 35, 468
476.
Gillespie, J.H.: 1981b, Evolution of the Mutation Rate at a Heteotic Locus, Proceedings of
the National Academy of Sciences (USA) 81, 80098013.
Goldenfeld, N. and Kadanoff, L.P.: 1999, Simple Lessons from Complexity, Science 284,
8789.
Holsinger, K.E. and Feldman, M.W.: 1983, Modifiers of Mutation Rate: Evolutionary
Optimum With Complete Selfing, Proceedings of the National Academy of Sciences
(USA) 80, 67326734.
Holsinger, K.E., Feldman, M.W. and Altenberg, L.: 1986, Selection for Increased Mutation
Rates with Fertility Differences Between Matings, Genetics 112, 909922.
Kauffman, S.: 1993, The Origins of Order, Oxford Unversity Press, Oxford.
Kauffman, S.: 1995, At Home in the Universe, Oxford University Press, Oxford.
Kitcher, P.: 1989, Explanatory Unification and the Causal Structure of the World, in P.
Kitcher and W. C. Salmon (eds), Scientific Explanation, University of Minnesota Press,
Minneapolis, pp. 410506.
Kitcher, P.: 1993, The Advancement of Science, Oxford University Press, Oxford.
Lande, R.: 1976, Natural Selection and Random Genetic Drift in Phenotypic Evolution,
Evolution 30, 314334.
Lenoir, T.: 1982, The Strategy of Life, D. Reidel, Dorcrecht.
Levins, R.: 1963, Theory of Fitness in a Heterogeneous Environment. II. Developmental
Flexibility and Niche Selection, American Naturalist 97, 7590.
Levins, R.: 1968, Evolution in a Changing Environment, Princeton University Press, Prin-
ceton.
Lewontin, R.C.: 1957, The Adaptation of Populations to Varying Environments, Cold Spring
Harbor Symposia on Quantitative Biology 22, 395408.
Lieberman, U. and Feldman, M.W.: 1985, Modifiers of Mutation Rate: A General Reduction
Principle, Theoretical Population Biology 30, 125142.
Maynard Smith, J.: 1976, Evolution and the Theory of Games, American Scientist 64, 4145.
Maynard Smith, J.: 1975, The Theory of Evolution, Penguin, London.
 683

Maynard Smith, J.: 1982, Evolution and the Theory of Games, Cambridge University Press,
Cambridge.
Mitton, J.B.: 1994, Molecular Approaches to Population Biology, Annual Review of Ecology
and Systematics 25, 4570.
Nicolis, G. and Prigogine, I.: 1993, Exploring Complexity, Freeman, New York.
Parrish, J.K. and Edelstein-Keshet, L.: 1999, Complexity, Pattern, and Evolutionary Trade-
Offs in Animal Aggregation, Science 284, 99101.
Richardson, R.C.: 1984, Biology and Ideology: The Interpenetration of Science and Values,
Philosophy of Science 51, 396420.
Richardson, R.C.: 1997, Natural and Artificial Complexity, Philosophy of Science 64
(Proceedings), S255S267.
Richardson, R.C. and Burian, R.M.: 1992, A Defense of Propensity Interpretations of
Fitness, in A. Fine, M. Forbes and K. Okruhlik (eds), PSA 1992, Volume 1, Philosophy
of Science Association, pp. 349362.
Roughgarden, J.: 1979, Theory of Population Genetics and Evolutionary Ecology: An
Introduction, MacMillan, New York.
Salmon, W.C.: 1989, Four Decades of Scientific Explanation, in P. Kitcher and W. C. Salmon
(eds), Scientific Explanation, University of Minnesota Press, Minneapolis, pp. 3252.
Schmalhausen, I.I.: 1949, Factors of Evolution, Blakiston Press, New York.
Sober, E.: 1984, The Nature of Selection, Bradford Books/M.I.T. Press, Cambridge.
Sober, E.: 1983, Equilibrium Explanation, Philosophical Studies 43, 201210.
Stearns, S.C.: 1989, The Evolutionary Significance of Phenotypic Plasticity, Bioscience 39,
436445.
Sultan, S.E. and Bazzaz, F.A.: 1993a, Phenotypic Plasticity in Polygonum persicaria. I.
Diversity and Uniformity in Genotypic Norms of Reaction to Light, Evolution 47,
10091031.
Sultan, S.E. and Bazzaz, F.A.: 1993b, Phenotypic Plasticity in Polygonum persicaria. II.
Norms of Reaction to Soil Moisture and the Maintenance of Genetic Diversity, Evolution
47, 10321049.
Sultan, S.E. and Bazzaz, F.A.: 1993c, Phenotypic Plasticity in Polygonum persicaria. III. The
Evolution of Ecological Breadth for Nutrient Requirement, Evolution 47, 10091031.
Wagner, A.: 1999, Causality in Complex Systems, Biology and Philosophy 14, 83101.
Weber, B.H.: 1998, Origins of Order in Dynamical Models, Biology and Philosophy 13,
133144.
Weber, B.H. and Depew, D.J.: 1996, Natural Selection and Self Organization, Biology and
Philosophy 11, 3365.
Wright, S.: 1931, Evolution in Mendelian Populations, Genetics 16, 97159.
Wright, S.: 1932, The Roles of Mutation, Inbreeding, Crossbreeding, and Selection,
Proceedings of the Sixth International Congress of Genetics 1, 356366.
Zawidzki, T.W.: 1998, Competing Models of Stability in Complex, Evolving Systems:
Kauffman vs. Simon, Biology and Philosophy 13, 541554.