The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe...

13
The Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ, 1,2,3 SERGI VALVERDE, 1,2 MARTI ROSAS CASALS, 1,4 STUART A. KAUFFMAN, 3 DOYNE FARMER, 3 AND NILES ELDREDGE 5 1 ICREA-Complex Systems Lab, Universitat Pompeu Fabra, Dr Aiguader 80, 08003 Barcelona, Spain; 2 Institut de Biologia Evolutiva, CSIC-UPF, Passeig Maritim de la Barceloneta, 37, 08003 Barcelona, Spain; 3 Santa Fe Institute, 1399 Hyde Park Road, Santa Fe New Mexico 87501; 4 Sustainability Measurement and Modeling Lab, Universitat Politecnica de Catalunya-Barcelona Tech, EET-Campus Terrassa, 08222, Barcelona, Spain; and 5 Division of Paleontology, American Museum of Natural History, New York 10024 Received 21 November 2012; accepted 19 December 2012 Technological evolution has been compared to biological evolution by many authors over the last two centuries. As a parallel experiment of innovation involving economic, historical, and social components, artifacts define a universe of evolving properties that displays episodes of diversification and extinction. Here, we critically review previous work comparing the two types of evolution. Like biological evolution, technological evolution is driven by descent with variation and selection, and includes tinkering, convergence, and contingency. At the same time, there are essential differences that make the two types of evolution quite distinct. Major distinctions are illustrated by current specific examples, including the evolution of cornets and the historical dynamics of information tech- nologies. Due to their fast and rich development, the later provide a unique opportunity to study technological evolution at all scales with unprecedented resolution. Despite the presence of patterns suggesting convergent trends between man-made systems end biological ones, they provide examples of planned design that have no equivalent with natural evolution. © 2013 Wiley Periodicals, Inc. Complexity 18: 15–27, 2013 Key Words: Technology; evolution; tinkering; convergence; information technology; culturomics 1. INTRODUCTION W e can look at the history of technology as a human- driven, parallel experiment of evolution [1–4]. So far artifacts are not capable of self-reproduction, but the Correspondence to: Ricard V. Solé, ICREA-Complex Systems Lab, Universitat Pompeu Fabra (GRIB), Dr Aiguader 80, 08003 Barcelona, Spain (e-mail: [email protected]) population-level dynamics of long-term technological inno- vation nonetheless resemble biological evolution in many ways. The design of new technologies is strongly influenced by existing technologies, and technological change can be viewed as a process of descent with variation and selection [5–7]. Both chance and the appropriate context are required for innovations to occur. Lineages of design often show rapid change and diversification as well as exaptation. The later is illustrated by Gutenberg’s printing press, when an existing © 2013 Wiley Periodicals, Inc., Vol. 18, No. 4 DOI 10.1002/cplx.21436 Published online 8 February 2013 in Wiley Online Library (wileyonlinelibrary.com) COMPLEXITY 15

Transcript of The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe...

Page 1: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

The Evolutionary Ecology of TechnologicalInnovations

RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2 MARTI ROSAS CASALS,1,4 STUART A. KAUFFMAN,3 DOYNE FARMER,3

AND NILES ELDREDGE5

1ICREA-Complex Systems Lab, Universitat Pompeu Fabra, Dr Aiguader 80, 08003 Barcelona, Spain; 2Institut deBiologia Evolutiva, CSIC-UPF, Passeig Maritim de la Barceloneta, 37, 08003 Barcelona, Spain; 3Santa Fe Institute,1399 Hyde Park Road, Santa Fe New Mexico 87501; 4Sustainability Measurement and Modeling Lab, Universitat

Politecnica de Catalunya-Barcelona Tech, EET-Campus Terrassa, 08222, Barcelona, Spain; and 5Division ofPaleontology, American Museum of Natural History, New York 10024

Received 21 November 2012; accepted 19 December 2012

Technological evolution has been compared to biological evolution by many authors over the last two centuries.As a parallel experiment of innovation involving economic, historical, and social components, artifacts define auniverse of evolving properties that displays episodes of diversification and extinction. Here, we critically reviewprevious work comparing the two types of evolution. Like biological evolution, technological evolution is drivenby descent with variation and selection, and includes tinkering, convergence, and contingency. At the same time,there are essential differences that make the two types of evolution quite distinct. Major distinctions are illustratedby current specific examples, including the evolution of cornets and the historical dynamics of information tech-nologies. Due to their fast and rich development, the later provide a unique opportunity to study technologicalevolution at all scales with unprecedented resolution. Despite the presence of patterns suggesting convergent trendsbetween man-made systems end biological ones, they provide examples of planned design that have no equivalentwith natural evolution. © 2013 Wiley Periodicals, Inc. Complexity 18: 15–27, 2013

Key Words: Technology; evolution; tinkering; convergence; information technology; culturomics

1. INTRODUCTION

W e can look at the history of technology as a human-driven, parallel experiment of evolution [1–4]. So farartifacts are not capable of self-reproduction, but the

Correspondence to: Ricard V. Solé, ICREA-Complex SystemsLab, Universitat Pompeu Fabra (GRIB), Dr Aiguader 80, 08003Barcelona, Spain (e-mail: [email protected])

population-level dynamics of long-term technological inno-vation nonetheless resemble biological evolution in manyways. The design of new technologies is strongly influencedby existing technologies, and technological change can beviewed as a process of descent with variation and selection[5–7]. Both chance and the appropriate context are requiredfor innovations to occur. Lineages of design often show rapidchange and diversification as well as exaptation. The later isillustrated by Gutenberg’s printing press, when an existing

© 2013 Wiley Periodicals, Inc., Vol. 18, No. 4DOI 10.1002/cplx.21436Published online 8 February 2013 in Wiley Online Library(wileyonlinelibrary.com)

C O M P L E X I T Y 15

Page 2: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

technology (the screw press) that was co-opted to servea completely novel purpose. Extinction and replacementare also common. As soon as a genuinely novel inventionappears, it is typically followed by an enormous diversifi-cation, followed by the extinction (and turnover) of mostcompeting inventions. Moreover, technological change alsodisplays convergence: Similar discoveries are made simulta-neously by different inventors [4, 8] such as the more than20 different patents involving light bulb inventions prior toEdison’s success [9]. The view that technological evolutionfollows similar rules to biological evolution has captured theinterest of scientists, historians, and engineers alike.

Despite the commonalities, technological evolutiondeparts from biological evolution in fundamental ways. Fortechnological change long-term goals and expectations playa leading role in which the designers seek optimality, typi-cally under explicit criteria such as efficiency, cost, and speed.Moreover, as pointed by François Jacob [10–13], in contrastto artifacts, living structures are largely the result of tinker-ing, that is a widespread reuse and combination of availableelements to build new structures.Technology is highly depen-dent on the combination of preexisting [2, 4] inventions,but unlike biology, the introduction of new simple elementscan completely reset the path of future technologies. In con-trast, in biology, once established, solutions to problems areseldom replaced.

Both biological and technological innovations involve costconstraints. Thermodynamic efficiency can also help under-standing the origin of some structures. Allometric scalinglaws provide a good illustration of how a theory of bio-logical distribution networks (including both vascular andrespiratory ones) based on efficient energy dissipation onfractal trees [14]. Efficiency has also been driving techno-logical improvements and marks the development of thesteam engine and the bicycle [15]. The evolution of the lattercan be traced as a succession of improvement steps towardsincreasing performance and lower metabolic cost. However,the coupling between energy costs and improvements isnot a precondition for technological change to occur. Onone hand, many examples illustrate a common pattern ofdevelopment of a given invention: In early stages, inventionsare often overly expensive and not perceived as economi-cally relevant. The barrier to diffusion can only be overcomethrough the vision of individuals pursuing their views andgoals [4].

Is it possible to formulate a theory of technological evo-lution? How much can we take advantage from our theoret-ical understanding of biological evolution? Recent advanceswithin network theory and a unique availability of the fos-sil record of human inventions might help in reaching thatgoal. Such theory needs to consider the existence of universaltrends, the economic context [2, 6, 16, 17] and history. Webelieve that a major effort in this direction would settle thedebate on similarities versus differences.

2. LOCK-IN AND CONTINGENCYA fingerprint of biological evolution is the existence of frozenaccidents in the spread and stabilization of a given trait.Similarly, economic constraints can enhance technologicallock-in [2, 6]. The QWERTY keyboard is an archetypal exam-ple, where a suboptimal configuration of keys was adopteddue to a conflict with typing speed. Even if rational evaluationof a technology and its alternatives reveals that the dominanttechnological solution is suboptimal, issues of compatibil-ity often make it impossible for better solutions to enter[2, 3, 7, 18]. Videocassette recorders, aircrafts, or clocks allprovide examples where choices were strongly affected byhistory. This can be explained theoretically by considering thepositive feedback associated to the imitation dynamics char-acteristic of technological adoption of novelties [19]. A sim-ple model illustrates the path-dependent nature of choicesmade in a system where preferential decisions are based ofmajorities.

Let us consider a population where users of a given tech-nology can choose between two possibilities A and B. Imaginethat we represent individuals choosing A and B as black andwhite squares on a lattice. At the beginning, both alterna-tives are equally chosen and we place A (black) and B (white)agents at random over the lattice. Assuming that agents fol-low the majority rule, a given agent will change its currentstate, say A → B, if the majority of its neighbors is adoptingthe B state. For convenience, we have projected our simula-tions on a sphere. As a consequence of the contagion process,domains of agents adopting the same choice emerge andexpand in time. If we start from a random lattice with A and Bagents, small fluctuations are amplified towards either all-Aor all-B equilibria (bottom spheres).

The lattice model can be easily formalized in mathemati-cal terms. If a and b indicate the relative fraction of individualschoosing each strategy. Encounters among individuals can bedescribed in a mean field approximation by using a simplemathematical model. Let a and b define the populations ofagents adopting A and B, respectively. Agents using A (B) willadopt technology B (A) provided that they perceive that someminimal number of individuals are using the alternative. Thedifferential equation for a reads:

dadt

= μa(η − b) − aφ(a, b) (1)

The right-hand side of the equation described two terms. Thefirst corresponds to an infection process: With an efficiency μ

and provided that the number of agents adopting B is belowsome threshold η. This can be interpreted as follows: There isa minimal number η of agents using B that define a criticalmass able to promote the shift A → B Similarly, an equationfor b is obtained as follows: db/dt = μb(η − a) − bφ(a, b). Ifa constant population constraint (a + b = 1) is assumed, itis easy to show that φ(a, b) = μ(η(a + b) − 2ab). The model

16 C O M P L E X I T Y © 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

Page 3: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

becomes one-dimensional and we can write an equation forthe time evolution of a:

dadt

= Fμ(a) = 2μa(1 − a)

(a − 1

2

)(2)

The term a(1 − a) is just a logistic function (with saturationwhen a = 1). Equation (2) actually corresponds to the popu-lation dynamics of one locus and two identical alleles due toselection alone [20]. The last term in the right-hand side cap-tures the majority rule. If a < 1/2, the A choice is in a minoritysituation and a decrease in A use will occur. Instead, when amajority has adopted A (namely a > 1/2) growth will occur.The previous equation has three equilibrium points, obtainedfrom da/dt = 0. These are a∗

0 = 0, a∗1 = 1, and a∗

2 = 1/2. Theirstability is obtained by the sign of

λμ(a∗i ) =

(∂Fμ(a)

∂a

)a=a∗

i

(3)

The first two correspond to winner-takes-all stable points,where A has failed to be adopted (a∗

0) and has invaded thewhole population. These are stable points, while the thirdone, associated to the coexistence between both technolo-gies, is unstable.

A useful analog of the path-dependent nature of thismodel is obtained by computing the associated potentialVμ(a) for the previous model [19]. The potential is linked tothe dynamical equation through

dadt

= − ∂

∂a

[−

∫Fμ(a)da

](4)

where the integral within the parenthesis is the potentialfunction. As defined, it will exhibit a maximum (minimum)when a∗ is unstable (stable). For our example, this gives:

Vμ(a) = −μ

2

(a4 − 2a3 + a2

)(5)

The curve Vμ(a) is displayed in Figure 1. The two wells sep-arated by a maximum are the accessible states and the sym-metric, unstable one (since 0 ≤ a ≤ 1, the values outsidethis domain are meaningless). The spheres shown as insetsare lattice simulations at different time steps (the sphericalprojection is used for convenience). Here, black and whiterepresent A and B states. As we can see, the control parame-ter μ only affects the depth of the alternative valleys but nottheir location. However, smaller μ values will be more likely tobe affected by stochastic fluctuations and thus the symmetrybreaking process might be more difficult (or even impossible)to occur.

3. THE GENOTYPE–PHENOTYPE MAPPING: TECHNOLOGYMEETS (SOMETIMES) BIOLOGY

The presence of suboptimal choices, as illustrated in theprevious section, reminds us that nonlinear dynamics can

FIGURE 1

Choosing between two given, initially equivalent choices can be repre-sented by means of a two-well function. The top of this curve definesan unstable state. A marble located on this top will rapidly roll downtowards on of the two possible valleys. Such choice is highly dependenton the small fluctuations happening at the beginning of the process.As a consequence, choices are highly path dependent: the originalsymmetry is broken.

dominate over rational decisions. This effectively limits theideal picture of man-made artifacts as optimal, purposefuldesigns. However, in some well-known cases, optimal solu-tions are achieved and maintained. This is illustrated by theevolution of the fork [21]. A sequence of steps knife → two-tines → three-tines → four-tines fork took place slowly andthe final design has not changed since it was adopted in theseventeenth century. Four tines provide a good compromisebetween an object that performs its functions properly: It pro-vides a broad surface, it fits in the mouth and can hold bothlarge and small pieces of food. A fitness function �(N ) couldbe measured using a population of users playing with differ-ent designs. Here, Nt is the number of tines and we would liketo know how Nt relates to �(Nt ). We can imagine a simpleexperiment where different people use (repeatedly) differ-ent forks to different conditions and provide some score foreach design. The average over different users provides an esti-mate for the Nt → �(Nt ) mapping. In this simple example, asingle-peak would be present at Nt = 4. If we add the cost offabrication to the fitness (so that performance is computedover the number of tines) the peak would be sharper.

Cost constraints can strongly limit the potential solutionsassociated to a given problem. As a consequence, optimaldesigns sometimes lead to remarkable matching between

© 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

C O M P L E X I T Y 17

Page 4: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

biological and man-made solutions. As an example, the samelaws associated to wiring very large scale integrated (VLSI)circuits under strong packing constraints [22] are followed bybrains [23]. Convergence in design principles has also beenreported from the study of brain circuits in the visual cortex,which obey the same basic architecture of parallel computervision [24]. The close relationship between integrated circuitsand neural systems is provided by the so-called Rent’s rule,which defines a power law in networks that exhibit hierarchi-cal modularity. Assuming that we partition the system intosubsystems of size N , the rule establishes that the number ofconnections C linking the subset with the rest of the systemscales as

C = 〈k〉N p (6)

where 〈k〉 and p are characteristic parameters. Here, 〈k〉 givesthe average number of links per node, whereas 0 ≤ α ≤ 1is the so-called Rent’s exponent. The later is typically lim-ited to the interval α ∈ [0.45, 0.75]. Given the importanceof the following result, it is worth to show how is this scalinglaw derived. Imagine a given system � (a network) involv-ing N0 elements (gates in a circuit, neurons, or groups ofneurons in a brain) and consider a given subsystem of size1 ≤ N ≤ N0. These components are connected forming anetwork. We indicate by 〈k〉 the average number of links pernode and we take a subset �k(N ) of elements within a givenarea, including N components. This subset will be connectedto the rest of the system � −�k by means of a number of linkscrossing its boundaries. Let us also indicate as C(N ) the num-ber of links connecting �k with the rest of the system. If weadd additional elements to �k , expanding it in size, new linksshould be expected to connect the larger subset with the restof the network. Indicating by C and N the increase in con-nectivity and size associated to the expansion, we can expecta relationship:

C = p(N )

(CN

)N (7)

The parameter p ∈ (0, 1) indicates a potential dependencybetween size and the degree of optimization involved in thecircuit design. VLSI designers use a number of strategies toreduce C and in particular a modular, hierarchical systemis specially For a random network where any element canconnect to any other one with the same probability, we wouldobserve p = 1. The last equation can be solved by using acontinuous approximation in terms of a differential equation,which gives:

C(N ) = 〈k〉 exp

[∫ N

1

p(N )

NdN

](8)

assuming that p(N ) = p, we obtain the standard formu-lation of Rent’s rule. In most known systems, this constantassumption seems the correct one.

FIGURE 2

Hierarchical modularity. In both circuits and neural systems, whichexperience cost constraints, a hierarchy of nested modules has beenshown to exist, displaying some seemingly universal regularities, suchas Rent’s rule. This rule establishes that the amount of connectionsC between elements in a subsystem of size N (here three scales areshown) with the rest of the system scales as C = 〈k 〉N p .

The fact that the exponent is homogeneous tells us that,

despite the scale change associated to increasing N , the rela-

tive fraction of links crossing boundaries scales with system’s

size in a constant way. This is characteristic of fractal objects

(where the basic pattern is repeated at different scales) and

thus an indication of the presence of hierarchical order.When

this method was applied to neural networks, a striking con-

vergence was found. Both the neural web of the nematode

C. elegans and human cortical maps shared a Rent’s expo-

nent p close to what is the expected value for an optimally

efficient hierarchical system. Such convergence, that shares

many properties in common with VLSI circuits, illustrates

the role played by cost constrains in promoting convergent

designs.

A single peak defines the simplest landscape for opti-

mization. But the presence of lock-in reminds us that fitness

landscapes are, in general, much more complicated. The

ruggedness of the landscape is connected to the nature of

the genotype–phenotype mapping [25–28]. The complexity of

this mapping pervades the path-dependent character of evo-

lution, how robust designs are and how evolvable they can be.

It is in general far from trivial to know the landscape structure

that pervades a given complex system. One exception is pro-

vided by those system where a discrete sequence space and

a well-defined phenotype space can be defined and scanned.

18 C O M P L E X I T Y © 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

Page 5: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

FIGURE 3

Neutral landscapes are common in a number of biological systems. They have been shown to drive the evolution of RNA and proteins through sequence space.In (a) we display an idealized picture of a small piece of an RNA landscape (Redrawn from [13]). Here, each node represents a given sequence, whereas linksamong nodes are possible paths connecting single-mutation events. Three different neutral networks are represented, associated to three different RNA folds.These landscapes are also present in technological systems. The analysis of field programmable gate arrays or FPGAs in short [31] has shown that a similarspace can be defined, with given arrangements of gates as nodes and connections indicating possible transitions from one circuit to another (see text). In (b)a rank plot distribution of these circuits is shown, using a small array (inset, Modified from Raman and Wagner, 2012).

A success story in this context is the analysis of the genotype–

phenotype mapping in both RNA and protein networks (refs).

Here, a mapping can be defined between a set of sequences

H(S) (nucleotide chains) and the set corresponding forms

H(F) resulting from the folding process:

: H(S) −→ H(F) (9)

Si −→ Fi (10)

It was shown that, starting from a given sequence Si ∈ S

it is possible to move through sequence space [Figure 3(a)]

through neutral mutations sharing the same phenotype Fi

(a common fold). Sequence space is enormously vast. If the

alphabet has size A and sequences have length L we have AL

possible sequences, an already hyperastronomic number. By

introducing single mutations starting from a given sequence,

it is possible to see if it has an impact on the shape (function)

or instead is neutral. A key finding was that we can travel

through this vast combinatorial space by spreading through

very large neutral networks of sequences folding in the same

shape (i.e., the same phenotype). In the simplified example

shown here, three sequences are indicated. Two of them (S1

and S2) span finite neutral networks, whereas S3 percolates

through sequence space.

The unexpected comes when we compare the previ-ous results with the corresponding mapping associated toa seemingly disparate object: electronic circuits. These arenow defined as sets of logic gates connected through a givenwiring scheme [29]. The circuit performs a function definingthe phenotype: For the given set of inputs (a fixed number)there is a specific output that defines the function performed.In other words, the wiring diagram is now the genotypeand the Boolean function implemented the phenotype. Thefull list of outputs describes the Boolean function �i , thatis, the phenotype P ≡ �i . The genotype is defined as theordered string of all weights wij . Two different sets, � andW describe the universe of possible functions and wirings,that is, the sets of possible phenotypes and genotypes. Theset of (genotype–phenotype) mappings between wiring andfunction is defined as:

: H(W) −→ H(�) (11)

where H(W) is the space of all possible FFNs and H(�) thecorresponding space of all Boolean functions. For each Wi ∈H(W), we have a genotype which maps into a given functiondescribed by a given �i ∈ H(�).

When the landscape associated to circuits performingcomputations [30, 31] was analyzed, a very similar patternwas found. Here, the genotype is identified with the network

© 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

C O M P L E X I T Y 19

Page 6: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

of elements and their connections, whereas the computa-tion being performed describes the function. By "amutating"the wiring diagram of a given circuit (altering single con-nections), it can be shown that neutral networks are alsowidely present, the boundaries of which are occupied by cir-cuits able to implement a great variety of novel functions. Acommon statistical regularity shared by both RNA and FPGAlandscapes is the broad distribution of abundances, shownin Figure 3(b). Here, we display the probability of finding agiven circuit, against its rank (see Ref. [31] for details). TheP(k) rank distribution provides the relative frequency of thekth most common circuit. The distribution is highly skewed,and follows a law, the Zipf’s law, namely:

p(k) = k−γ∑nj=1 j−γ

(12)

with γ ≈ 1. This is a highly skewed distribution, with a fewsequences (or circuits) being highly common whereas mostare rare.

A major consequence of remarkably common pattern isthat key structural properties of the organization of man-made designs can have an unexpected potential for evolvabil-ity [31]. What these results do indicate [29–31] is that designprinciples in engineering could benefit from understandingevolution on landscapes.

4. DIVERSIFICATION, EXTINCTION, AND TECHNOLOGICALREGIMES

The rise of innovations often leads to both the decline of pre-existing solutions and the creation of new ones. An exampleis provided by the evolution of cornets [32, 38] which diversi-fied over a century until valve trumpets replaced them, aftera radiation in styles and usages, somewhat similar to bioticreplacements. Similarly, the invention of the car marked thedecline and fall of carriages, harness shops and buggies, whilefavoring the expansion of the oil industry. Paved roads, gasstations, or motels came later. Although a given innovationoccurs in a well-defined economic context, novel technolo-gies can deeply transform how the economy is organized andhow new economic regimes emerge. Events like the rise ofelectronics and personal computers represent a turning pointfrom which a new technological regime is born [34–37]. Suchregimes can be described in terms of the grammar or rule-set embedded in the way a given technology must be used.They include many facets such as product standards anddesign and engineering practices. The resetting of a regimeand its replacement by a new one share some similaritieswith recovery from mass extinction [2, 3, 38].

Properly tracing the rise and fall of technologies is notalways possible. As always in science, the availability of data isan important determinant of what we know, and the data forbiological versus technological evolution are very different.For many technologies, we have good anecdotal evidence on

performance and prevalence through time. Data on perfor-mance can take many forms. For example, for informationtechnologies we have reasonably good historical data forproperties such as processing speed, information storage, ortransmission, both in terms of cost and physical character-istics [39–41]. There is no corresponding historical perfor-mance data in biology—we can only guess at the strength orrunning speed of dinosaurs.

In contrast, in biology we have a well-developed phy-logeny and rich data on the history of phyla with no tech-nological equivalent. This is not an accident—as discussedlater, the phylogeny of technologies is not hierarchical, butrather is more similar to that of bacteria. While many studieshave been done for specific families of technologies, there areno comprehensive databases on technological change. As aresult, the data we have and the inferences that we can makeabout biological versus technological change are very differ-ent. Two major sources of quantitative analysis helps fillingthis gap.

5. TEMPORAL TRENDS AND LEARNING CURVESOne of the biggest differences between technology and biol-ogy is that for technologies we have clear notions of progress.For example, in 1936 Theodore Wright conjectured that theunit cost of a given model of airplane drops as a well-definedpower law in the cumulative number produced. This “law”has since been observed, at varying degrees of fidelity, for alarge number of other technologies [39–44], and the generalpattern goes under the name of an experience or learningcurve. An alternative hypothesis about the rate of techno-logical improvement is due to Gordon Moore, who proposedthat the density of transistors for integrated circuits increasesexponentially with a doubling time of about two years [45].The rule of exponential change also holds quite well for manyother aspects of information technology. An approximateexponential relationship between unit cost and other mea-sures as a function of time has recently been shown to holdfor many other technologies as well [39, 41].

Moore’s law and Wright’s law are closely related: If pro-duction increases exponentially and unit cost drops expo-nentially, then Wright’s law holds. Scaling laws are commonlyobserved between cost C and production z, i. e.

C(z) ∼ z−α (13)

where the scaling exponent α weights the ratio of improve-ment. Some general trends displayed by learning curvesare consistent with expected trends displayed by popula-tions climbing rugged fitness landscapes [43, 46–49]. In thiscontext, design structure matrices (DSMs) offer a promisinglink between technological evolution and its underlying fit-ness landscape [50, 51]. As illustrated in Figure 4, DSM aredescribed by an n × n matrix M = (Mij) with Mij �= 0 if thecomponent i has some cost impact on component j (and is

20 C O M P L E X I T Y © 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

Page 7: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

FIGURE 4

Example design structure matrices (DSMs) with n 14 13 components.Black squares represent links. The DSM on the left was randomly gen-erated to have fixed out-degree for each component. The DSM on theright represents the design of an automobile brake system. All diago-nal elements are present because a component always affects its owncost.

zero otherwise).They capture the internal constraints derivedfrom the interaction of different components, providing arepresentation of the genotype–phenotype mapping and howit is affected by epistatic interactions among units and howthey affect scaling laws.

A simple example illustrates the approach followed inRef. [51]. Imagine an idealized scenario where a given tech-nology has a cost c(t) which can change over time as newinnovations occur. For simplicity, we consider a finite rangeof values, to be normalized with uniform distribution f (c) on[0, 1]. In other words, we have

∫ 1

0f (c)dc = 1 (14)

Cost will be reduced if chance events allow a new design to beused at a smaller cost cÕ. It can be shown that the equationdescribing the change of cost follows:

dcdt

= −〈c〉2

P[c′ < c] (15)

= −〈c〉2

∫ c

0f (c)dc (16)

= − 12

〈c〉2 (17)

which can be easily solved, giving a power law decay:

c(t) = c(0)

1 + c(0)t/2≈ 1

t(18)

where c(0) indicates initial cost and the last term is theapproximation for long times. As we can see, this simplemodel provides a correct prediction for the time decay of costin technological evolution. More detailed models can be builtby considering the internal structure of the technological arti-fact and how different parts affect others performance [51].By using DSMs, an effective mapping between the network

of hardware interactions and the function being performedis defined. Because of this, the DSM captures pleiotropiceffects and how they influence the rates of improvement andthe existence of internal limitations. It has been shown [51]that the matrix connectivity pattern (the density of links orthe presence of modularity) largely determines the speed ofimprovement, which can be continuous or punctuated.

6. INNOVATION NETWORKSThe assumption that combinatorics is at the heart of innova-tion provides a possible rationale for the accelerating growthof innovations. If pairs of components can be combined withsome probability, the number of new designs will increasewith the number of already present designs. This picturehas led to the suggestion [52] that technological improve-ment leads inevitably to singularities, with the number ofinventions exploding on a finite time scale.

The idea that innovations happen mainly by combination[2] predicts (assuming unlimited resources) that the num-ber of inventions will grow in time following a hyperboliccurve [53, 54]. Specifically, if x indicates the number of inven-tions and we assume that combining previous designs createsnew inventions, the time evolution of x would follow, in itssimplest form, the equation

dxdt

= μ(t)x2 (19)

which says that a reaction between two given inventions isneeded to create new ones. Here, μ(t) is the rate at whichtwo inventions are used (by an inventor or group of inven-tors) to generate new ones. In general, this would be a time-dependent parameter, since it might depend on the numberof inventors, economic factors or aging phenomena. Themodel can be easily solved, leading to

x(t) = x(0)

1 − x(0)∫ t

0 μ(τ)dτ(20)

where x(0) would be the initial number of inventions. Forthe simplest scenario where μ can be considered constant, asingularity exists at a finite time

τ = 1μx(0)

(21)

when x becomes infinite.One consequence of such pattern is that a system where

combination is easy to happen will rapidly explore the adja-cent possible [4, 55] defining the next layer of potential designsobtainable through combination of existing ones. Due to theimmense potential for combination, the diversity of potentialnovel inventions could make the outcome essentially unpre-dictable [55]. However, even when very large search spacesneed to be explored, solutions are nevertheless found. This

© 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

C O M P L E X I T Y 21

Page 8: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

is illustrated by drug search in chemical space [56] of whichonly an extremely small fraction is screened. This suggeststhat the combinatorial potential might be strongly reduceddue to the underlying syntax constraining what objects canbe linked together.

Despite the appeal of the previous picture, dedicatedefforts using different measurements and databases suggesta more complex scenario [39]. Let us take for example thenumber of patents [Figure 1(a)] over time. The inset showsa transition from the industrial revolution to the so-calledsecond industrial revolution and characterized by the rise ofmass production and the rapid growth of chemical, electri-cal, and automotive industries and the availability of energyresources. Similarly, the global record of citations has a broadaccelerated trend of growth, particularly obvious in the lastdecades, strongly influenced by the rise of software (coloredarea) which started in the 1960s and rapidly accelerated dur-ing the 1980s. As a written description of innovations, patentsprovide a surrogate of technological dynamics [57–59]. Therapid growth in patent diversity have been compared toPhanerozoic diversity time series [3] but once again the sim-ilarity is only partially significant. During periods of crisis orwar their numbers fall. This is shown in [Figure 5(a)], wherewe indicate the starting point of only five of them using blackdots. Here, we pointed: (1) American Civil War, (2) USA enterWorld War I, (3) great depression, and (4) USA enters WorldWar II. However, a different type of external trigger can alsoprovoke a decay in patenting, as shown by Eq. (5) which indi-cates introduction of strict intellectual property legislationand higher costs for patenting rights. Such top-down controlhas obviously no biological counterpart.

Patent inventories give us a kinetic view of innovation,but much more can be uncovered by looking at how patentsare related to each other. Since a given patent must citeothers from which the novel technology gets inspiration, anetwork linking a given patent with past ones can be defined.The number of outward links starting from a patent relatesit to previous inventions. Similarly, the number of incom-ing links provides a rough estimate of its importance. Usingavailable data sets, a network of patent citations can be con-structed [59] [Figure 5(b)]. Each node represents a patent andarrows indicate that a citation is made to a previous invention.The central node in this case is the oldest invention. Colorsindicate modules: Groups of patents that have more con-nections among them than with others within the network.These modules correlate with groups of related inventions,suggesting that there is considerable structure in the way newtechnologies emerge.

Patent networks are highly heterogeneous, with somehubs involving many citations, whereas most nodes have onlyone or two [59]. This pattern can be accounted for using arich-gets-richer mechanism favoring highly connected nodes(the popular patents) [59–62]. By following the time evolu-tion of citations, it is actually possible to detect two major

FIGURE 5

Patent citations and innovation by combination. In (a) we display thetime series of USA patents from 1837 to 2009. Drops in citation recordsare often associated to crashes and wars 1–4) but also due to legislation(5).A closer picture is obtained by looking at the actual pattern of cross-citations among patents. This defines a network (b) where nodes arefiled patents and the (directed) links are citations among them. Thenetwork shown here corresponds to early inventions in tomography(see Ref. [59]) and colors indicate topological modules, that is, subsetsof patents that share a larger number of citations among themselvescompared with the rest of the dataset. Topological modules are usuallyrelated to functional subsets with specialized traits.

components affecting the relative importance of an inven-tion. The first is the rapid growth of citations often associatedwith the combinatorial potential of the invention. The secondis an aging process: as patents get older, they tend to be lesscited. This reflects the changing landscape of technology andthe inevitable replacement of old by new designs.

Can phylogenetic analyses help untangling the originsof technological change? Unfortunately, the differences areagain important. As noted in Ref. [33] the post-industrialpace of innovation is highly labyrinthine in terms of how

22 C O M P L E X I T Y © 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

Page 9: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

FIGURE 6

Lateral transfer muddies phylogeny in material cultural evolution, as illustrated here by the evolutionary tree of cornets. Vertical lines represent the periods ofmanufacture of distinct cornet models, here plotted against time. Red arcs indicate so called reticulations, that is, potential nonvertical instances of informationtransfer among makers. Here, two main groups are indicated, corresponding to the so called Stölzel (left rectangle) and Périnet (unmarked right domain)cornets (Redrawn after Ref. [32]).

information is exchanged between inventions and their mak-ers. It is now accepted that an appropriate approach tocultural evolution requires a network perspective: trees areclearly not enough. The dichotomous picture of branchingis being replaced by reticulate phylogenies (see below). InFigure 6, the problem is illustrated with the reconstructed his-torical evolution of cornets [32] where the vertical branchescorrespond to periods of manufacture of particular models.Instruments equiped with so-called Stölzel and Périnet valvesystems, respectively. Nonvertical links (indicated by curvedred lines) are reticulations, that is, instances of informationtransfer among makers. The open circles at the bottom markkey innovations in cornet design, including: (1) valve num-ber, (2) shifting of the second valve slide and valve alignment,(3) changing of bell exit position and bell placement, and(4) alteration of bell shape (trumpetization). The reticulatednature of the diagram is largely due to the rapid and largeinformation exchange between makers, who introduced dif-ferent types of innovations and changes in the number andlocation of valves and other features. The data set providesclear illustrations of gradual evolution, stasis followed by

punctuation, extinction, and selection, but also examplesof a special feature of cultural evolution: the Lazarus effect.In contrast to living species, artifacts (such as the so-calledsheperd’s crook design) can "resurrect" based on availableinformation about them.

7. EVOLVING INFORMATION TECHNOLOGYSoftware systems have a relatively recent role within techno-logical evolution and are currently the most widespread andinfluential systems in our society and economy. They wereintroduced in the 1950s (Figure 7, colored areas) as soon ascomputer hardware and a theory on computing architecturesemerged [65]. Hardware was clearly a precondition for soft-ware to emerge, as it might have been in early life [66] but,once in place, a rapid coevolution unfolded. In spite of itsrelevance, software has been largely ignored in the literatureon technological evolution. In many ways, software allows usto get closer to the picture of biological evolution where agenomic description is present, since software deals mainlywith performing functions and is coded [8, 67].

© 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

C O M P L E X I T Y 23

Page 10: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

FIGURE 7

The software hierarchy. At the largest scale involving the actual origins of programming languages (a) we can appreciate a pattern of diversification andmultiple fusion that reveals not a tree but a network of innovations. Here, we highlight the different branches associated to special classes of languages(yellow and violet represent procedural languages, blue depicts object-oriented, green indicates an important class of languages used in artificial intelligence,and red represents system programming languages). Multiple branches can fuse on a single, novel programming language (such as Javascript) through aplanned process of large-scale integration of different components.Within each language, large-scale projects can be described in terms of a network (b,Aztecsoftware) connecting different subsets of functionally well-defined blocks (b). Here, each node in the graph for one of these large-scale programs representsone such functional block, which depend (are linked to) on others. A directed arrow thus indicates that a given block needs to "call" another block. Suchnetwork is scale free and modular, being dominated by a handful of hubs and having most elements just one or two links with others. Such a structure seemsto be a consequence of the widespread copy-paste process of growth. The rules of reuse also explain the biased statistical presence of some special, smallsubgraphs known as network motifs. In (c), we show some of the most common network motifs found in the graph displayed in (b). The frequency of thesemotifs has been shown to be a byproduct of a duplication-rewiring process [63, 64] thus not (necessarily) related to special advantages of given structuresat the small scale.

One of the greatest advantages of software is that, on allscales, its fossil record is extremely well preserved, involv-ing multiple levels of detail, from written code to the socialnetworks of engineers that designed it. At a coarse grainedscale, a phylogeny of programming languages can be defined.Some of them are displayed in Figure 7(a). After the inven-tion of the first programming language, new versions cameup to facilitate the interactions between man and machine.In this context, software could run on different hardwarethus making it much more adaptive [67]. FORTRAN in par-ticular became a standard for engineering and physics andstill survives, mainly because its simplicity and its first-move character: Libraries developed to solve general calcu-lations are constantly reused. Many other languages failedand got extinct [68]. The diagram reveals both radiationsand merging. We can appreciate this in the branching from

Algol-60, which solved some syntax problems associatedto FORTRAN and was as much hardware-independent andportable as possible.

Later on, other major transitions took place, includingthe emergence of so-called Object-Oriented Programmingwhere instructions and data are both included within well-defined structures called objects [67, 68]. This implies a mod-ularization of the program organization that enhances theirevolvability. As it happens with cornets, engineers were ableto merge different parts of previous systems and incorpo-rate them into new languages. But here the scale of suchevents is much larger and conceptually more complex. Thecarefully planned merging of different structures used indifferent languages and their correct integration have nomatch in any other evolutionary trend. In this respect, theevolution of programming languages deeply differentiates

24 C O M P L E X I T Y © 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

Page 11: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

from any biological counterpart, including the evolution oflanguage.

Beyond the language programming scale, a hierarchy oflevels of software evolution is defined by the long-term devel-opment of complex networks of interacting pieces of codeand data. Although software engineering is obviously anintentional, purpose-driven activity, it is strongly affectedand shaped by reuse.The evolution of large programs hasbeen shown to occur in parallel to an ever-increasing fre-quency of copy-paste [67–70]. An unintended consequenceof this tinkering is that, as it occurs with patent networks,it spontaneously generates complex modular graphs [Figure7(b)]. Their patterns of organization seem to be explainedthrough extensive cloning [58, 59] thus questioning Jacob’sseparation between technology and biological evolution interms of the presence or absence of reuse. Even at the small-est scale the statistical pattern of network motifs, defined assmall subgraphs whose frequency strongly departs from ran-dom expectations can be accounted for by reuse [Figure 2(c)].As it occurs with other systems [12, 71–74] a nonadaptiveinterpretation of their biased frequency is obtained.

8. DISCUSSIONIt seems undeniable that technological change displaysnumerous life-like features, suggesting a deep connectionwith biological evolution. This is specially well illustrated bycommon convergent patterns, particularly when optimiza-tion due to cost/energy constraints is at stake. But some"disanalogies" [75] are also noticeable. The coexistence ofboth similarities and dissimilarities occurs because techno-logical and biological change share universal traits and causalmechanisms that pervade their long-term trends, no mat-ter how different they appear at a lower scale. What is at theroot of these potential universals? Statistical physics has beensuccessful in showing that disparate systems often behave inthe same way (once they are properly scaled) and that thisuniversal behavior can be explained by considering only avery reduced number of properties [76–78]. It has been found,for example, that a magnet, a flock of birds or even the stockmarket behave similarly in terms of displaying tipping points,transient dynamics, or response to external stresses. Thisuniversal behavior is present in the structure of neutral land-scapes, the existence of frozen accidents or the punctuatedequilibria.

Any theory of technological change will require newmetaphors and tools, particularly when dealing with infor-mation systems. Reticulate networks, instead of trees [79–81]provide the adequate test to the role played by pure combi-nation. Networks allow defining modules and their relationsand the success within language evolution [82, 83] indicatesthat they are more appropriate when dealing with culturaldynamics. Additionally, technological design can be obtainedthrough in silico Darwinian selection [84–87] thus defining anew level of complexity. As the distance between technology

FIGURE 8

A pure engineering design can be obtained by following a Darwinianselection process.Using computer simulations,NASAAmes researchersevolved a population of antennas with a fitness function defined interms of a signal-processing problem. The best fit candidates wererather organic in their shape, as the one shown here. Image from NASAAmes Center

and biology shrinks, some unexpected results emerge ashardware is itself evolved.

One important message is that the results of Darwinianselection often escape from our intuition and expertise, sug-gesting that part of the design landscape is not accessibleto our minds. This is illustrated by hardware design involv-ing Darwinian selection. Genetic algorithms in particular[73–75] allow a population-level exploration of vast parame-ter spaces. In this approach, each member of the populationis more or less fit in solving the given problem, and its survivaland number of offspring directly depends on its functionalsuccess. Surviving solutions are copied with mutations: smallchanges are introduced with some given probability.

Evolved circuits, engines, game strategies, or market fore-casts can be obtained by using evolutionary programming.Often these algorithms provide improved modifications thatmake the previous designs more efficient. However, somechallenge our intuition. An example is shown in Figure 8,where an evolved antenna is displayed. It was obtained underevolutionary dynamics to solve a difficult task on a NASAmission satellite to study the Earth magnetosphere. Using apopulation of candidate structures and evolving them overmany generations, each antenna was tested against a numberof technical requirements that were simulated on a super-computer. The evolved structure is far from what an engineerwould ever had imagined. This example offers a very interest-ing situation where the peaks in the technological landscapethat solve the problem cannot be captured by the mentalstructure of the human designer.

Another differential trait of technology is the role playedby social and economic factors [88] on innovation. Networks

© 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

C O M P L E X I T Y 25

Page 12: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

of inventors act as ecosystem engineers [89] able to transformtheir own innovation landscape (see Ref. 90) thus chang-ing the ways other technologies will interact. Moreover, asearly noticed by Robert Solow, economic growth requiresthe presence of innovation [91] and thus coevolution amongeconomy and technology must be included. Similarly, weneed to identify the scales at which technological hierar-chies operate. In biology, such hierarchy can be described[92] including different levels, from population dynamics togenotype–phenotypic maps.

In technology, population changes can be estimated usingculturomics [93], that is, exploiting the abundance of givenstrings of words (the so called n-grams) extracted from mil-lions of digitalized books. The relative frequency of thesen-grams within the culturome provides a measure of theirimportance over time. A kinetic theory of n-grams could

take advantage of useful ecological metaphors used withinother domains of cultural evolution, such as language change[94]. Defining appropriate mappings between genotype andphenotype is a much more difficult task [95]. Once again,information technology, with all its richness and multiplicityof scales, offers our best bet to achieve this goal. There is awhole fossil record there to be unearthed.

ACKNOWLEDGMENTSThe authors thank Peter Schuster and Ricardo Haussmanfor useful comments. This work has been supported byresearch grant FIS2004-05422 (RS,SV), by the James McDon-nell Foundation (RS), the Marcelino Botin Foundation (SV,RS), the TEKES Foundation of Finland (SK), the National Sci-ence Foundation (NSF) Grant SBE0738187 (DF), and by theSanta Fe institute.

REFERENCES1. Basalla, G. The Evolution of Technology, Cambridge University Press, 1989.2. Arthur, B. The Nature of Technology. What it is and How it Evolves; Free Press, 2009.3. Kelly, K. What Technology Wants; Viking, 2010.4. Johnson, S. Where Good Ideas Come From: A Natural History of Innovation; Riverhead Press, 2010.5. Sahal, D. Patterns of Technological Innovation; Addison-Wesley, 1981.6. Schuster, P. Untamable curiosity, innovation, discovery and bricolage. Complexity 2006, 11, 9–11.7. Arthur, B. Increasing Returns and Path Dependence in the Economy; Michigan University Press, 1997.8. Ogburn, W.F.; Thomas, D. Are inventions inevitable? A note on social evolution. Polit Sci Quart 1922, 37, 83–98.9. Friedel, R.; Israel, P. Edison’s Electric Light: Biography of an Invention; Rutgers Univercity Press, 1986.

10. Jacob, F. Evolution as tinkering. Science 1977, 196, 1161–1166.11. Wilkins, A. Between design and bricolage: Genetic networks, levels of selection and adaptive evolution. Proc Natl Acad Sci USA 2007, 105, 8590–8596.12. Solé, R.V.; Ferrer, R.; Montoya, J.M.; Valverde, S. Selection, tinkering, and emergence in complex networks. Complexity 2002, 8, 20–33.13. Schuster, P. The origin and processing of biological information. In: Front. Life 2001, 1, 329–346.14. Brown, J.H.; West, G.B. Scaling in Biology; Oxford University Press, 2000.15. Minetti, A.E.; Pinkerton, J.; Zamparo, P. From bipedalism to bicyclism: Evolution in energetics and biomechanics of historic bicycles. Proc R Soc Lond B 2001,

268, 1351–1360.16. Farmer, D. Market force, ecology and evolution. Ind Corp Change 1999, 11, 895–953.17. Eldredge, N. Evolution in the marketplace. Struct Change Econ Dyn 1997, 8, 385–398.18. Gould, S.J. The PandaÕs thumb of technology. Nat Hist 1987, 96, 14–23.19. Solé, R.V. Phase Transitions; Princeton University Press, 2011.20. Kimura, M. Diffusion models in population genetics. J Appl Prob 1964, 1, 177–232.21. Petrovski, H. The Evolution of Useful Things; Vintage Books, 1992.22. Christie, P. The interpretation and application of Rent’s rule. IEEE Trans VLSI Syst 2000, 8, 639–648.23. Bassett, D.S.; Greenfield, D.L.; Meyer-Lindenberg, A.; Weinberger, D.R.; Moore, S.W.; Bullmore, E.T. Efficient physical embedding of topologically complex

information processing networks in brains and computer circuits. PLoS Comp Biol 2010, 6, e1000748.24. Nelson, M.E.; Bower, J.M. Brain maps and parallel computers. Trends Neurosci 1990, 13, 403–408.25. Schuster, P.; Fontana, W.; Stadler, P.F.; Hofacker, I.L. From sequences to shapes and back: A case study in RNA secondary structures. Proc R Soc Lond B 1994,

255, 279–284.26. Huynen, M.A.; Stadler, P.F.; Fontana, W. Smoothness within ruggedness: The role of neutrality in adaptation. Proc Natl Acad Sci USA 1996, 93, 397–401.27. Nimwegen, E.; Crutchfield, J.P.; Huynen, M. Neutral evolution of mutational robustness. Proc Natl Acad Sci USA 1999, 96, 9716–9720.28. Wilke, C.O. Adaptive evolution on neutral networks. Bull Math Biol 2001, 63, 715–730.29. Ferrer, R.; Janssen, C.; Solé, R.V. Topology of technology graphs: Small world patterns in electronic circuits. Phys Rev E 2001, 64, 32767.30. Macia, J.; Solé, R.V. Distributed robustness in cellular networks: Insights from synthetic evolved circuits. J R Soc Interface 2009, 6, 393–400.31. Raman, K.; Wagner, A. The evolvability of programmable hardware. J R Soc Interface 2012, 55, 269–281.32. Tëmkin, I.; Eldredge, N. Phylogenetics and material cultural evolution. Curr Antrop 2007, 48, 146–153.33. Eldredge, N. Paleontology and cornets: Thoughts of material cultural evolution. Evo Edu Outreach 2011, 4, 364–373.34. Nelson, R.R.; Winter, S.G. An Evolutionary Theory of Economic Change; Belknap Press, 1982.35. Nelson, R.R.; Winter, S.G. In search of a useful theory of innovation. Res Policy 1977, 6, 36–76.36. Rip, A.; Kemp, R. Technological change. In: Human Choice and Climate Change, Vol. 2; Rainer, S.; Malone, L., Eds.; Batelle Press, 1988; pp. 327–400.37. Van den Ende, J.; Kemp, R. Technological transformations in history: How the computer regime grew out of existing computing regimes. Res Policy 1999, 28,

833–851.38. Arthur, B.; Polak, W. The evolution of technology within a simple computer model. Complexity 2006, 11, 23–31.

26 C O M P L E X I T Y © 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

Page 13: The evolutionary ecology of technological …complex.upf.es/~sergi/Sole_Complexity_EvoTech.pdfThe Evolutionary Ecology of Technological Innovations RICARD V. SOLÉ,1,2,3 SERGI VALVERDE,1,2

39. Nagi, B.; Farmer, J.D.; Trancik, J.E.; Gonzales, J.P. Superexponential long-term trends in information technology. J Tech Forecast Soc Change 2011, 73,1061–1083.

40. Koh, H.; Magee, C.L. A functional approach for studying technological progress: Application to information technology. Tech. Forecast. Soc. Change 2006, 73,1061–1083.

41. Farmer, J.D.; Trancik, J. Dynamics of technological development in the energy sector. In: London Accord Final Publication; Onstwedder, J.P.; Mainelli, M., Eds.,2007; pp. 1–24.

42. Argote, L.; Epple, E. Learning curves in manufacturing. Science 1990, 247, 920–924.43. Auerswald, P.; Kauffman, S.A.; Lobo, J.; Shell, K. The production recipes approach to modeling technological innovation: An application to learning by doing.

J Econ Dyn Control 2000, 24, 389–450.44. McDonald, A.; Schrattenholzer, L. Learning rates for energy technologies. Energ Policy 2001, 29, 255–261.45. Moore, G.E. Cramming more components onto integrated circuits. Electronics 1965, 38, 114–117.46. Kauffman, S.A.; Levin, S. Towards a general theory of adaptive walks on rugged landscapes. J Theor Biol 1987, 128, 11–45.47. Kauffman, S.A. At Home in the Universe. Chapter 9; Oxford University Press., 1995.48. Frenken, K. A complexity approach to innovation networks. The case of the aircraft industry. Res Policy 2000, 29, 257–272.49. Marengo, L.; Valente, M. Industry dynamics in complex products spaces: An evolutionary model. Struct Change Econ Dyn 2010, 21, 5–16.50. Baldwin, C.Y.; Clarck, K.B. Design Rules; MIT Press, 2000.51. McNerney, J.; Farmer, J.D.; Redner, S.; Trancik, J.E. Role of design complexity in technology improvement. Proc Natl Acad Sci USA 2011, 108, 9008–9013.52. Kurtzweil, R. The Singularity is Near; Viking, 2005.53. Hanel, R.; Kauffman, S.A.; Thurner S. Phase transitions in random catalytic networks. Phys Rev E 2005, 72, 036117.54. Hanel, R.; Kauffman, S.A.; Thurner, S. Towards a physics of evolution: Critical diversity dynamics at the edges of collapse and bursts of diversification. Phys

Rev E 2007, 76, 036110.55. Kauffman, S.A. Investigations, Oxford University Press, 2000.56. Lipinski, C.; Hopkins, A. Navigating chemical space for biology and medicine. Nature 2004, 432, 855–861.57. Tuomi, I. Networks of Innovation; Oxford University Press, 2006.58. Jaffe, A.B.; Trajtenberg, M. Patents, Citations and Innovations; MIT Press, 2003.59. Valverde, S.; Solé, R.V.; Bedau M.; Packard, N. Topology and evolution of technology innovation networks. Phys Rev E 2007, 76, 032767.60. Gress, B. Properties of the USPTO patent citation network: 1963–2002. World Patent Info 2010, 32, 3–21.61. Bommarito II, M.J.; Katz, D.M.; Zelner, J.L.; Fowler, J.H. Distance measures for dynamic citation networks. Physica A 2010, 389, 4201–4208.62. Verspagen, B. Mapping technological trajectories as complex networks. Adv Complex Syst 2009, 10, 93–115.63. Valverde, S.; Ferrer, R.; Solé, R.V. Scale-free networks from optimal design. Europhys Lett 2002, 60, 512–518.64. Valverde, S.; Solé, R.V. Network motifs in computational networks: A case study in software architecture. Phys Rev E 2005, 72, 026107.65. Williams, M.R. A History of Computer Architecture; Prentice-Hall, 1985.66. Dyson G Darwin Among the Machines; Addison-Wesley, 1997.67. Sammet JE Programming Languages: History and Fundamentals; Prentice-Hall, 1969.68. Mens T.; Demeyer, S Software Evolution; Springer-Verlag, 2008.69. Myers, C.R. Software systems as complex networks: Structure, function, and evolvability of software collaboration. graphs Phys Rev E 2003, 68, 048116.70. Concas, G.; Marchesi, M.; Pinna, S.; Serra, N. Power-laws in a large object-oriented software system. IEEE Trans Soft Eng 2007, 33, 687–708.71. Solé, R.V.; Valverde, S. Are network motifs the spandrels of cellular complexity? Trends Ecol Evol 2006, 21, 419–22.72. Solé, R.V.; Valverde, S. Spontaneous emergence of modularity in cellular networks. J Roy Soc Interface 2007, 5, 129–133.73. Lynch, M. The evolution of genetic networks by non-adaptive processes. Nat Rev Genet 2007, 8, 803–813.74. Wagner, G.; Pavlicev, M.; Cheverud, J.M. The road to modularity. Nat Rev Genet 2007, 8, 921–931.75. Ziman, J. Technological Innovation as an Evolutionary Process, Cambridge University Press, 2000.76. Kauffman, S.A. The Origins of Order: Self-Organization and Selection in Evolution; Oxford University Press, 1993.77. Cowan, G.A.; Pines, D.; Meltzer, D. Complexity: Metaphors, Models and Reality; Perseus Books, 1996.78. Mitchell, M. Complexity: A Guided Tour; Oxford University Press, 2009.79. Makarenkov, V.; Legendre, P. From phylogenetic tree to a reticulated network. J Comput Biol 2004, 11, 195–212.80. Lipo, C.P.; O’Brien, M.J.; Collard, M.; Shennan, S.J. Mapping our Ancestors; Aldine Transaction Publishers, 2009.81. Kunin, V.; Goldovsky, L.; Darzentas, N.; Ouzounis, C.A. The net of life: Reconstructing the microbial phylogenetic network. Genome Res. 2005, 15, 954–959.82. Dagan, T.; Martin, W. Getting a better picture of microbial evolution en route to a network of genomes. Phil Trans R Soc Lond B Biol Sci 2009, 364, 2187–96.83. Nelson-Sathi, S.; List, J.-M.; Geisler, H.; Fangerau, H.; Gray, R.D.; Martin, W.; Dagan, T. Networks uncover hidden lexical borrowing in Indo-European language

evolution. Proc Royal Soc London B 2011, 278, 1794–1803.84. Mitchell, M. An Introduction to Genetic Algorithms; Bradford Books, 1998.85. Forrest, S. Genetic algorithms: Principles of natural selection applied to computation. Science 1993, 261, 872–878.86. Koza, J.R.; Keane, M.A.; Streeter, M.J. Evolving inventions. Sci Am F 2003, 882, 52–59.87. Long, J. Darwin’s Devices; Basic Books, 2012.88. Jones, C.G.; Lawton, J.H.; Shachak, M. Organisms as ecosystem engineers. Oikos 1994, 69, 373–386.89. Erwin, D.H. Macroevolution of ecosystem engineering, niche construction and diversity. Trends Ecol Evol 2008, 23, 304–310.90. Hidalgo, C.A.; Klinger B.; Barabási, A.-L.; Hausmann, R. The product space conditions the development of nations. Science 2007, 317, 482–487.91. Solow, R. A contribution to the theory of economic growth. Q J Econ 1952, 76, 65–94.92. Schuster, P. How does complexity arise in evolution?. Complexity 1996, 2, 22–30.93. Michel, J.-B.; Shen, Y.K.; Aiden, A.P.; Veres, A.; Gray, M.K.; Pickett, J.P.; Hoiberg, D.; Clancy, D.; Norvig, P.; Orwant, J.; Pinker, S.; Nowak, M.A.; Aiden, E.L.

Quantitative analysis of culture using millions of digitized books. Science 2011, 33, 176–182.94. Solé, R.V.; Corominas-Murtra, B.; Fortuny, J. Diversity, competition, extinction: The ecophysics of language change. J Royal Soc Interface 2010, 7, 1647–1664.95. Lobo, J.; Miller, J.H.; Fontana, W. Neutrality in technological landscapes; Santa Fe Institute Working Paper, 2004.

© 2013 Wiley Periodicals, Inc.DOI 10.1002/cplx

C O M P L E X I T Y 27