Complex Adaptive Systems Author(s): John H. Holland Reviewed work(s): Source: Daedalus, Vol. 121, No.

1, A New Era in Computation (Winter, 1992), pp. 17-30 Published by: The MIT Press on behalf of American Academy of Arts & Sciences Stable URL: http://www.jstor.org/stable/20025416 . Accessed: 27/10/2012 17:57
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The MIT Press and American Academy of Arts & Sciences are collaborating with JSTOR to digitize, preserve and extend access to Daedalus.

http://www.jstor.org

John H. Holland

Complex Adaptive Systems

One

of

the

most

important

roles

to act as a simulator air over an

of physical it provides

a computer can play is a com When processes. a unique of

puter mimics the behavior of a system, such as the flow of

airplane wing, us with way

studying the factors that control that behavior. The key, of course, is
to offer an accurate of the system being rendition for the computer have scored some major In the studied. past fifty years, computers successes in this regard. Designers of airplanes, bridges, and Ameri to their designs ca's Cup yachts all use computers routinely analyze we to metal. For such artifacts, them know how before they commit to mimic a century the behavior quite exactly, using equations interest discovered over ago. there are systems of crucial However, so far defied accurate simulation have

to humankind

that

Economies, by computer. and the brain all immune systems, developing embryos, ecologies, at compre that block broadly based attempts exhibit complexities For example, the equation-based methods that work well hension. A have a much more limited scope for economies. for airplanes finance in asking the accuracy a an as to out can of the computer impact engineer policy change play to play out the implications in asking the computer of tilting expect an airplane wing. minister expect the cannot same

Despite the disparities and the difficulties, we are entering a new

era in our ability to understand come from two for optimism begun to extract a common and foster recent kernel from such systems. First, these advances. The grounds scientists have each of the

systems:

is Professor and Professor of Psychology John H. Holland at the University and Maxwell Engineering of Michigan Professor

Science and of Computer at the Santa Fe Institute.

17

18
systems

JohnH. Holland
involves a similar

structure." That is, these systems "evolving to their component parts to adapt themselves change and reorganize reason the problems their is main This the the posed by surroundings. are difficult to understand a and control?they constitute systems are "moving target." We that mediate these systems vations are would suggest. These that learning, are much that the mechanisms however, more alike than surface obser and the deeper are now similarities under that a

mechanisms the

systems important enough grouped common name, complex systems. adaptive is the new era in computation The second relevant advance are not computer

is

the theme of this issue of Dcedalus. This advance will allow experts
who complex examine code. adaptive a model same of particular models savvy to "flight-test" a can directly For maker systems. example, policy for its "reality," without the underlying knowing policy maker can then formulate and try out

That

different policies on themodel, again without becoming involved in the underlying coding, thereby developing an informed intuition
about future effects It is the thesis models, when of the policies. of this article that around offer these new computation-based structural kernel opportunity: vis-?-vis

constructed

the common a much-needed and useful and

of

systems, adaptive complex enable the formulation of new problems COMPLEX To arrive ranging from

They

policies

trade balances

sustainability

major to AIDS.

ADAPTIVE

SYSTEMS of complex systems?to adaptive and what makes them complex

at a deeper understanding what makes them understand

to look at a particular is useful the system. Consider adaptive?it It consists immune of of numbers system. units, large highly mobile that continually called antibodies, repel or destroy an ever-changing cast of invaders Because called antigens. (bacteria and biochemicals), the invaders could all come in an almost infinite variety of forms, the immune

system cannot simply develop a list of all possible invaders. Even if it

that take the time to do so, there is simply not room enough to store or information. the immune Instead, system must change that has made these systems so hard to simulate.

adapt ("fit to") its antibodies as new invaders appear. It is this ability
to adapt

Complex Adaptive Systems

The immune system faces the additional complication

19

that itmust

distinguish self from other; the systemmust distinguish the legitimate

from parts of its owner task because herculean cast of invaders. This is a the ever-changing the owner's cells and their biochemical

constituents number in the tens of thousands of kinds. Mistakes in identification do occur in some individuals, giving rise to the usually
fatal autoimmune so good diseases, at self-identification its own the that but they are rare. The immune it provides that, at present, those in a skin graft system is our best for of of

scientificmeans of defining individuality. An immune systemwill not

even confuse example. How adaptation does cells with from a sibling, process levels

immune enables

the ongoing system manage it to achieve such remarkable

identification?We do not really know, though there are interesting conjectures with varying degrees of evidence.Models of this complex adaptive system are hard to formulate. It is particularly difficult to
provide different experts in the area with models that allow "thought exper

iments," models
mechanisms

that enable the expert to develop intuition about

and organizations.

We
adaptive

face similar problems when dealing with the other complex

systems.1 All of a kaleidoscopic

them involve of parts great numbers of simultaneous interactions. array undergoing They all seem to share three characteristics: evolution, behavior, aggregate and anticipation. As time goes on, the parts evolve in Darwinian fashion, attempting

to improve the ability of their kind to survive in their interactions with the surrounding parts. This ability of the parts to adapt or learn
Some systems. complex adaptive systems are quite simple: a thermostat adaptive adapts by turning the at a constant furnace on or off in an attempt to keep its surroundings the of interest here are However, processes temperature. adaptive is the pivotal characteristic of

complex because they involve many parts and widely varying indi
vidual criteria thermostat) Complex to the constant temperature (analogous a "good outcome" for what would be. systems also exhibit an aggregate sought behavior by the that

adaptive

is not simply derived from the actions of the parts. For the immune system this aggregate behavior is its ability to distinguish self from
other. For an economy, it can range from the GNP to the overall

network of supply and demand; for an ecology, it is usually taken to

20

JohnH. Holland

be the overall food web or the patterns of flow of energy and materials; for an embryo, it is the overall structure of the developing individual; for the brain, it is the overt behavior it evokes and controls. Generally, it is this aggregate behavior thatwe would like to
and modify. To do so, we must understand how the from the interactions behavior of the parts. emerges aggregate As if this were not complex there is a further feature that enough, In seeking makes these systems still more complex?they anticipate. can to adapt to changing the parts be thought of as circumstance, understand developing rules that anticipate the consequences of certain re

sponses. At the simplest level, this is not much different from Pavlovian conditioning: "If the bell rings, then food will appear."
However, even for simple conditioning, the effects are quite complex

when large numbers of parts are being conditioned in different ways. This is particularly the case when the various conditionings depend
the resulting antici between parts. Moreover, even when cause in aggregate behavior, major pation changes an come even if it not true. of oil do The anticipation shortage, they never comes to pass, can cause a sharp rise in oil prices, and a sharp upon can increase ability complex There in attempts to anticipate energy sources. This emergent is one of the features we least understand about one most is it of the systems, yet important. to find alternative the interactions

adaptive technical that needs is one final, more point, emphasis. are Because the individual system parts of a complex adaptive their rules for each interaction, ("conditioned") revising continually

part is embedded in perpetually novel surroundings (the changing

behavior defined of the other for the system economics, on optimal get there." parts). As a result, as a whole. the aggregate behavior of the

system is usually far from optimal, if indeed optimality can even be

in physics, concentrate tems "never and elsewhere, For this reason, standard theories are of little help because they whereas sys end-points, complex adaptive to continue and evolve, they steadily They

exhibit new forms of emergent behavior. History and context play a critical role, further complicating the task for theory and experiment.
at a some parts of the system may settle down temporarily Though or are if "dead" local optimum, uninteresting they usually they It is the process of for an extended period. remain at that equilibrium that we must end points, rather than the never-reached becoming, we are to if gain insight. study

Complex Adaptive Systems

MASSIVELY PARALLEL COMPUTERS

21

The introduction of the digital programmed computer profoundly

changed Massively thousands our view of what parallel of interconnected be accomplished made computers?computers could with computation. of up of hundreds

microcomputers?will produce changes It is not just a matter of speed, though that that are equally profound. a massively can handle Because is important. parallel computer large new it of actions numbers offers ways of displaying simultaneously, and interacting with data. systems adaptive that workstation's as far beyond It provides of studying ways complex as the reach of a current workstation

are beyond the reach of an adding capacities a or machine slide rule. Indeed, massively should parallel computers a in the of revolution sys investigation produce complex adaptive to revolution tems comparable the introduction of the produced by in biology.2 microscope The effects longer-range of massive parallelism are not easy to

predict at this early stage, but a little hindsight offers some clues. At the beginning of the computer era, in the 1940s and early 1950s,
most with ability as they unfolded, the of those increases magnitude problems. us. They made in with decreases amazed precipitous price, coupled word electronic mail, the personal processing, possible widespread and related sets of activities, work such as personal video station, But games economy people. and This simulations. This both and has altered process scientists computer an ever-increasing foresaw increasing to tackle speed and storage, along scientific and business

new major sectors of the has produced the work and play of large numbers of increases in speed and

storage, for massively by decreasing prices, is already underway accompanied The new "microscope" will soon be as pervasive parallel machines. as the personal workstation is today. MODELS OF COMPLEX ADAPTIVE SYSTEMS

of headlong

or rule, A complex system has no single governing adaptive equation, that controls the system. Instead, it has many distributed, interacting in the way of a central control. Each of little or nothing parts, with the parts participate is governed rules. Each of these rules may by its own an outcome, in influencing and each may influence the

22
actions

JohnH. Holland
becomes to adapt of these

structure of other parts. The resulting rule-based for that enable the system the evolutionary grist procedures a better to its surroundings.3 We can develop understanding structure. rule-based distributed, can into Most be parsed rules

evolutionary procedures ifwe first take a closer look at this idea of a

simple condition/action rules: If

[condition true], then execute [action]. The simplest rules in this form look much like specifications for psychological reflexes: If [the surface feels hot], then execute [a backward jerk of the hand]; if [there is a rapidly moving object in peripheral vision], then execute [a movement of the eyes until the object is in the center of the visual
sent by other rules, in rules act on messages field]. More complicated turn sending out their own messages: If [there is a message X], then execute activities can of message Y]. Quite complicated [transmission out by combinations of such rules; in fact, any computa in a computer that can be specified language can be carried out an of condition/action rules. combination by appropriate This distributed, many-ruled organization places strong requirements on computer simulation of complex adaptive systems. The most direct be carried tion approach a simulation in which many is to provide rules are active "natural" for massively parallel computation. simultaneously?a a distributed, rules can be active When many simultaneously,

novel

a system can handle perpetual novelty. On encountering a as car of road the side the with flat such "red tire," situation, by several relevant rules, such as those for "red," the system activates "car," "flat tire," and so on. It builds a "picture" of the situation from rule-based than The treating it as a monolithic is similar to whole that obtained never when before one advantage a face in terms of component whole. eyebrows,

parts rather encountered. describes an hair,

indecomposable forehead,

parts, rather than treating it as for the face? Select, say, 8 components and chin. nose, mouth, eyes, cheekbones,

Allow 10 variants for each component part?different hair colors and = textures, different forehead shapes, and so on. Then 108 100,000,000 faces can be described by combining these components in different ways. This at the cost of storing only 8 x 10 = 80 "building
block" similar components. combinations. Moreover, when blocks a building block is useful in one

combination, it is at least plausible that itwill prove useful in other,

Building for transferring previous experience to new thus give the system situations. a capacity

Complex Adaptive
Massive parallelism

Systems

23

a com in simulating is clearly an advantage terms in of of conceived system simultaneously acting plex adaptive can be allocated to each rule, while rules. An individual processor the connections between the processors provide for rule interactions. The

resulting model

is both natural and rapidly processed.

To provide the system must have ways of changing for adaptation, its rules. Such procedures give the system its characteristic "evolving are two structure." There kinds of computational that are procedures relevant: Credit credit assignment and rule discovery procedures procedures. one wants is because the system, and necessary assignment to evolve toward something. first requires Credit assignment

is, then it requires a way to pick "good" performance out and "reward" those parts of the system that seem to be causing A system that rewards good performance may good performance. never become but it can get better and better. optimal,

its rules, a sense of what

Credit assignment is a traditional problem in artificial intelligence

research. individual In a rule-based rules the object is to assign credit to system, to their contribution to the system's in proportion

overall (aggregate) performance. We

strength to the rule: The more

can think of this credit as a

a rule contributes to good assigned the stronger it becomes, and vice versa.4 By "stronger" performance, we mean that the rule, based on its past successes, is given a stronger voice in future decisions. As successive situations are encountered, the relevant rules compete to control behavior, the stronger rules being

the likelywinners. That is, if a rule has produced a good outcome in some situation in the past, then it ismore likely to be used in similar
situations Credit evolve rules. in the future. assignment to deal with this can enable new a system to select the best from the

rules it has, but it cannot supply the systemwith new rules. If it is to

the system will have to create new some kind of rule system requires discovery a is subtle Rule it is important process, because procedure. discovery that the discovery process generate plausible rules, rules that are not situations, For the

obviously wrong on the basis of past experience. The philosopher

C. S. Pierce reasoning commentary, on this matter.5 is quite informative To apply Pierce's to this model, it is convenient to think of rules as made up

of smaller pieces, or building blocks. My

then, is that the discovery

own version of Pierce's

and recombination of

24

JohnH. Holland
invented rules.6

building blocks is an important step toward assuring the plausibility

of newly

To approach rule discovery in terms of building blocks, it is useful

as "parents," parents. The process farmer to think of "breeding" and new rules. That is, strong rules are selected rules are produced the by crossing offspring have rules valuable is that strong assumption building strong

blocks inside them that should be incorporated into new rules. This
a breeder crosses horses or a the process whereby new varieties of hybrid corn. Here the object is to produces new hypotheses. to plausible rules that amount offspring produce mimics

Rule discovery procedures of this kind are called genetic algorithms.7 A genetic algorithm "learns" automatically by biasing future gener
ations blocks of rules toward (as, in genetics, in successive frequently in each of of above-average building ever more sets of genes appear coadapted It can be proved that genetic generations). combinations the known systems. Of complex adaptive are this nonsense, process by and are system "good" behavior

algorithms find and recombine useful building blocks. They have

counterparts course, many but nonsense atically of the new rules generated

rules do not promote out. weeded

This rule discovery procedure, once again, lends itself tomassively

is a simple operation strong parents Crossing computation. parallel on Because the computer. low processing that imposes requirements set of rules can be treated as a population, with mating the whole on is the parallelism population, simultaneously throughout going easily exploited. THE FUNDAMENTAL SYSTEMS ATTRIBUTE OF

INTERNAL MODELS: COMPLEX ADAPTIVE

There is still one property of complex adaptive systems thatwe have

to examine internal more models form and use systems adaptive closely. Complex on to anticipate the future, basing current actions

expected outcomes.8 It is this attribute that distinguishes complex adaptive systems from other kinds of complex systems; it is also this attribute that makes the emergent behavior of complex adaptive
to understand. intricate and difficult systems to note that we rarely think of anticipation, It is interesting as a characteristic in general, of organisms though prediction, or we

Complex Adaptive Systems

25

readily ascribe it to humans. Still, a bacterium moves in the direction of a chemical gradient, implicitly predicting that food lies in that direction. The butterfly thatmimics the foul-tastingMonarch butter
fly survives discourages because it implicitly A wolf predators. forecasts bases its actions that a certain wing pattern on anticipations

landmarks and scents. by a mental map that incorporates generated simulations itself represents man's The science of computer attempt

to make predictions ranging from the flight characteristics of yet untried aircraft to future GNP, but we have only recently been able to endow programs themselves with model-building capabilities. It is important that we understand the way inwhich complex adaptive systems build and use internal models, because so much of their
behavior stems from anticipations based on these internal models.

An internal model allows a system to look ahead to the future

consequences those actions. of current itself to actions, without actually committing can acts set In particular, the system avoid that would

it irretrievably down some road to future disaster ("stepping off a cliff"). Less dramatically, but equally important, the model enables
moves current "stage-setting" that set up later the agent to make moves The very essence of attaining that are obviously advantageous. a competitive in chess or economics, it is the whether be advantage, moves. execution of and stage-setting discovery An internal model whenever may, of course, be incorrect in some or many

ways. But then hindsight can be used to improve themodel; themodel

outcome fail to match subsequent can thus make system (credit assignment again). improvements overt rewards or detailed without information about errors. This is a ismodified its predictions The tremendous corrective advantage information one the rewards inmost occur is playing situations, where rewards or at end of long sequences the of only a game of chess or making a long-term real-world action are usually much delayed.

action. Whether investment,

for current

Internalmodels enable improvement in the interim. AN INTERIMSUMMARY Here's a condensed view of the description of complex adaptive
so far. The systems' basic components systems presented as sets of rules. The systems rely on three key mechanisms: and recombination. Parallelism ism, competition, permits are treated parallel the system

26
to use describe

JohnH. Holland
individual and act upon to marshal sets of rules to rules as building blocks, activating the changing situations. allows Competition its rules as the situation a torrent The of demands, providing it

the system

flexibility and transfer of experience. This is vital in realistic environ

the agent receives ments, where to current decisions. irrelevant credit assignment and rule most of information, for procedures adaptation? useful, repeatable

discovery?extract

events from this torrent, incorporating them as new building blocks. Recombination plays a key role in the discovery process, generating
plausible new, new rules contexts. these mechanisms extant allow a complex adaptive to respond, instant system balances system by exploration to from parts of tested rules. It implements the

heuristic that building blocks useful in the past will prove useful in
similar Overall, adapt, while to

using its environment.

capabilities In so doing the

instant,

(acquisition of new information and capabilities) with exploitation (the efficient use of information and capabilities already available). The system that results iswell founded in computational terms, and it does indeed get better at attaining goals in a perpetually novel
environment.

ACCESS

TO SIMULATIONS

OF COMPLEX

ADAPTIVE

SYSTEMS

Simulations produce "batch uncover

floods

of complex adaptive of data. The result on

processing"

on computers, executed systems, is reminiscent of the early days of as When the output appears computers:

interminable pages of printout and numerical tables, it is difficult to

or surprising much less react to interactions, significant can to user rather than experiment be reduced them. The observing,

ing and controlling. This need not be. we are tomake parallel simulations of complex adaptive systems If accessible, two criteriamust be satisfied. First, the parallel simulation must directly mimic the ongoing parallel interactions of the complex
adaptive interface user be a visual, there must Second, system.9 game-like not used to controls for experts natural that provides a For example, via computers. systems policy maker to try out an economic way, without model in much the way that a of the

exploring should be able made

pilot tries out a flight simulator. Actions and decisions should be

in the usual requiring any cognizance

Complex Adaptive
behaves in realistic ways in well-known situations.

Systems
This has

27

underlying computations. It should also be easy to see if the model

the

additional value of allowing experts to feed back "reality checks" to the simulation designers. Research initiatives at the Santa Fe Institute, in cooperation with a commercial firm, SimLabs, lead us to believe that powerful interfaces of this kind are possible for complex
adaptive CURRENT systems. SIMULATIONS OF COMPLEX ADAPTIVE SYSTEMS

We
kind

are only in the earliest stages of developing simulations of the

there are some and Sargent Fe results. The work suggestive on the evolution of money Institute.10 This study uses a

but just discussed, of Marimon, McGrattan, economics rule-based program agents

provides an early example. Itwas

at the Santa in a classical It shows

initiated in 1989 as part of the

combination of theory and simulation to study the effect of adaptive,

Wicksell's from economics, trading model even that when the artificial agents start triangle. soon a with decide medium of rules, they upon randomly generated and reach other exchange close-to-optimal trading patterns. Among to understanding the immune system studies, there is a new approach using a massively parallel computer,11 and an actual policy study

using data from the office of management and budget inMilan, Italy.12The latter is directly concerned with increasing the efficiency of decision making in the 730 offices scattered throughout the
Lombardy to discover it attained, was region. The study's major objective, which a were which from relevant to very large number, factors,

the various decisions made

information, These the director

by the local offices. By using this

decision procedures that would

structured

lead to increased efficiency in the local offices.

to the computer savvy, early results are really only accessible In all three of the models but they point the way. cited, the study of in the system's the mechanisms providing evolutionary changes structure will encourage more more accessible models. We realistic, can then expect current tial advances available research to expand exploratory to a wide range of users. into substan

MATHEMATICSAND THEORY Complex adaptive systems are so intricate that there is little hope of
a coherent theory without the controlled experiments that a mas

28

JohnH. Holland
the same time, in an area an theoretical experiments unguided by appropriate amount to more little than the pot usually "watching computer makes possible. At outside the experimental physics After all, no system currently under to inform theory, but without

sively parallel this complex, framework

boil." Sustained progress outside the guidelines of a theory is as

as attempting modern unlikely of theoretical physics. framework adaptive system. We need

investigation in physics is as complex as a full-fledged complex

experiments

theory all is lost.

Fortunately, mathematics we at which can bring there are several points to bear on the approach can show above. We outlined certain conditions, credit proce assignment appropriate

can the relevant rules. We strengthen stage-setting a that recombination, mediated does by genetic algorithm, of rules toward bias the population the use of above progressively are blocks.13 There also formal frameworks that average building also show to the process of generating internal models, with accompany some that establish of their On a ing proofs elementary properties.14 are there of mathematics from broader relevant pieces perspective, apply mathematical The economics and mathematical ecology that can be

that, under dures do indeed

generalized to apply to all complex adaptive systems.15

into a theory, a is to weld these disparate pieces challenge the of the processes pervasiveness theory that explains evolutionary the common kernel of all complex systems. The forming adaptive that assure the emergence the mechanisms of theory should elucidate internal models. Coordinated computer that simulations offer new should provide for

critical tests of the unfolding theory. The simulations should also

suggest tions well-informed inductions conjectures against directions

theory. The broadest hope is that the theoretician, by testing deduc

and the simulations, can reincarnate the cycle

of theory and experiment so fruitful in physics. To my knowledge there is only one organization, the Santa Fe Institute, that has taken the general mathematical study of complex
systems as its central mission.16 adaptive a unique its campus range of experts The institute has drawn to and in physics, economics,

related mathematical
with the University strengths particular

disciplines. It has formed a working

alliance

of Michigan

to take advantage of that university's in psychology, and business adminis sociology,

tration. Even though the institute is only five years old, these

Complex Adaptive
interactions of complex have already produced substantial

Systems

29

changes

in the study

adaptive

systems.

SUMMARY Complex adaptive systems represent the kernel of some of our most

difficult problems, ranging from trade balances to control of the AIDS epidemic. They can be simulated on massively parallel com
a network of interacting rule-based components. by defining natural interfaces such sim for By providing "flight-simulator-like" can we to these open ulations, systems exploration by policy makers not to and other experts who do have the time become computer puters

savvy. This has the double value of giving the designers "reality checks," while allowing policy makers to explore the differences effected by different policies. By looking for pervasive phenomena in
can implement we the classic hypothesize-test experiments, revise cycle for the study of complex systems. The experi adaptive such

mental part of this cycle is particularly important, because such

far from equilibrium, systems typically operate continually undergo to not revisions and do classic, ing improvements. They yield on mathematical that equilibrium-based approaches rely linearity,

attractors, fixed points, and the like. A new kind of mathematical

framework Without is required, one that emphasizes continuing adaptation

through recombination of building blocks.

such a framework, the computer-based will experiments an than uncoordinated into forays endlessly complex we can greatly domain. With such a framework, expand our under difficult questions. standing of these important, be little more

ACKNOWLEDGEMENTS Research by the author, relevant to this article, was supported in part by the National Science Foundation, under grant IRI-8904203, and by the Santa Fe Institute. Iwould like to thank James Bailey of Thinking Machines Corporation for many helpful suggestions after a careful reading of the manuscript.

30

lohnH. Holland

ENDNOTES ^ee D. L. Stein, ed., Lectures in the Sciences of Complexity (Reading, Mass.: Addison-Wesley, 1989); and P.W. Anderson, K. A. Arrow, and D. Pines, eds., The Mass.: Addison-Wesley, 1988). Economy as an Evolving Complex System (Reading, 2J.H. Holland, "A Universal Computer Capable of Executing an Arbitrary Number of Sub-Programs Simultaneously," Proc. 1959, Eastern Joint Computer Confer ence (New York: Institute of Electrical and Electronic Engineering, 1959), 108-13; andW. D. Hillis, The Connection Machine (Cambridge:MIT Press, 1985). 3See J. H. Holland, K. J. Holyoak, R. E. Nisbett, Processes of Inference, Learning, and Discovery
?Ibid., 70-75.

and P. R. Thagard, Induction: (Cambridge:MIT Press, 1989).

5C. S. Pierce, Collected Papers, Science and Philosophy, (Cambridge: Harvard University Press, 1958).

vol. 7, ed. A. W. Burks

Natural and Artificial Systems (AnnArbor: University H. Holland, Adaptation in 6J. of Michigan Press, 1975). 7Ibid., 89-140.
8J. H. Holland, "Emergent Models," in A. Scott, ed., Frontiers in Science (Cam

bridge: Blackwell,
9S. Forrest, "Emergent

1990).
Computation," in S. Forret, ed., Emergent Computation

(Amsterdam: Elsevier [North-Holland],

10R. Marimon, E. McGrattan, and T.

1990).
Money as a Medium of Exchange

J. Sargent,

in an Economy with Artificially Intelligent Agents, 89-004 (Santa Fe: Santa Fe Institute, 1989).

Santa Fe Working

Paper

11S. Forrest and A. Perelson "Genetic Algorithms and the Immune System," in H. Schwefel and R. Maenner, eds., Parallel Problem Solving from Nature (Berlin: Springer Verlag, 1991), 320-25. 12N.H. Packard, "Genetic Learning Algorithm Complex Systems 4 (1990): 543-72. 13Holland, Adaptation inNatural for the Analysis of Complex Data"

and Artificial Systems, chap. 7.

14Holland, Holyoak, Nisbett,

15Anderson, Arrow, and Pines,

and Thagard, Induction, appendices 2A and 2B.

eds., The Economy as an Evolving Complex System,

75-97,

205-41.

16D. Pines, ed., Emerging Syntheses in Science (Reading, Mass.: Addison-Wesley, 1987).