barandarian, x. moreno, a. (2008) adaptivity. from metabolism to behavior.pdf

http://adb.sagepub.com/

Adaptive Behavior

http://adb.sagepub.com/content/16/5/325The online version of this article can be found at:

DOI: 10.1177/1059712308093868

2008 16: 325Adaptive BehaviorXabier Barandiaran and Alvaro Moreno

Adaptivity: From Metabolism to Behavior

Published by:

http://www.sagepublications.com

On behalf of:

International Society of Adaptive Behavior

can be found at:Adaptive BehaviorAdditional services and information for

http://adb.sagepub.com/cgi/alertsEmail Alerts:

http://adb.sagepub.com/subscriptionsSubscriptions:

http://www.sagepub.com/journalsReprints.navReprints:

http://www.sagepub.com/journalsPermissions.navPermissions:

http://adb.sagepub.com/content/16/5/325.refs.htmlCitations:

at Ulle/Biblioteca Edif. Central on October 14, 2010adb.sagepub.comDownloaded from


http://adb.sagepub.com/content/16/5/325

http://www.sagepublications.com

http://www.isab.org.uk

http://adb.sagepub.com/cgi/alerts

http://adb.sagepub.com/subscriptions

http://www.sagepub.com/journalsReprints.nav

http://www.sagepub.com/journalsPermissions.nav

http://adb.sagepub.com/content/16/5/325.refs.html


325

Adaptivity: From Metabolism to Behavior

Xabier Barandiaran, Alvaro MorenoDepartment of Logic and Philosophy of Science, University of the Basque Country, Spain

In this article, we propose some fundamental requirements for the appearance of adaptivity. We

argue that a basic metabolic organization, taken in its minimal sense, may provide the conceptualframework for naturalizing the origin of teleology and normative functionality as it appears in living

systems. However, adaptivity also requires the emergence of a regulatory subsystem, which implies a

certain form of dynamic decoupling within a globally integrated, autonomous system. Thus, we ana-lyze several forms of minimal adaptivity, including the special case of motility. We go on to explain

how an open-ended complexity growth of motility-based adaptive agency, namely, behavior, requires

the appearance of the nervous system. Finally, we discuss some implications of these ideas forembodied robotics.

Keywords naturalist approach to normativity · autonomous systems · adaptivity · minimal agency ·

decoupling of the nervous system · definition of adaptive behavior

1 Introduction

Adaptivity is the capacity that certain systems possessto modify themselves in order to adjust to changes inthe environment. As is apparent in many types of sys-tem, this capacity is obviously an essential property ofliving beings that exhibit not only an enormous func-tional plasticity at the phylogenetic scale, but also a richrepertoire of responses to environmental changes insomatic time. However, many ecological and socialorganizations are also considered adaptive. Last but notleast, a certain number of artificial systems—from vir-tual computer networks to physical “autonomous”robots—are also designed to show adaptive capacities.

This generic idea of adaptivity obscures twoimportant philosophical and scientific questions: thepresupposed distinction between system and environ-ment and the fact that the idea of “adjustment” (andconsequently of adaptivity) implies an irreducible nor-mative dimension.

The first problem we call the “problem of iden-tity;” that is, from the set of possible and arbitrary sep-arations between system and environment, which arethe ones we choose and why. If we are to attributeadaptive capacities to a system, we must first specifywho is going to be adjusted to what. In addition, whentalking about adaptivity, we are not just referring to amere structural adjustment (a liquid adjusts to a con-tainer) between a system and its environment, but toa functional one; this brings us to the second prob-lem.

The second problem is the “problem of normativ-ity.” Normativity refers to the value attribution that isgiven to a process or object (e.g., adaptive or maladap-tive to an interaction or structure in an organism, trueor false to a cognitive state or belief, beautiful or uglyto a work of art, etc.). Normativity challenges physi-calist scientific approaches to the understanding of ourworld because it introduces a value asymmetry (good/bad, true/false, adapted/maladapted) in the description

Copyright © 2008 International Society for Adaptive Behavior(2008), Vol 16(5): 325–344.DOI: 10.1177/1059712308093868

Correspondence to: Xabier Barandiaran, Department of Logic and Philosophy of Science, University of the Basque Country, PO Box 1249, 20080 San Sebastian-Donostia, Spain. E-mail: [email protected].: +34 943015516; Fax: +34 943311056



326 Adaptive Behavior 16(5)

of nature. However, although alien to fundamentalphysics, normativity is an essential component ofbiology; whether a structure or interaction is adaptiveor maladaptive for an organism is a value judgmentthat a scientist engaged in the analysis and synthesisof adaptive systems must make.

According to these considerations, adaptivity is acapacity that a system possesses to maintain its iden-tity by compensating, through the use of certain inter-nal mechanisms, those environmental changes thatjeopardize its identity. However, a system that triggersfunctional interactions to compensate for dangerousenvironmental perturbations only because it is exter-nally guided to do so (such as, for example, a tele-operated robot) will never be a truly adaptive system(as the adaptivity lies in the intelligent system that oper-ates it). Thus, a truly adaptive capacity can only occurin systems that define their own goals and their ownnormative adjustment to the satisfaction of thesegoals. It is precisely in living systems where the proc-esses triggered are not divorced from their internalorganization; they are causally connected to its main-tenance. Thus, living systems are true adaptive agentsbecause they generate actions in order to achievegoals that they generate themselves.

Thus, a full understanding of adaptivity is stronglyrelated to the naturalization of concepts such as goal,normativity, identity, and functionality. This obliges usto look at the basic organization of living systems.However, even the simplest organism is a highly com-plex system. If we try to understand what adaptivity isfrom a naturalistic perspective, we have to search for theminimal organization that might support autonomous,plastic interactions with its environment.1 Explainingadaptivity requires us to specify under which basic orminimal organizational conditions it appears, giventhe best available law-like understanding of the uni-verse and the biological constraints that are observedupon them. Although it is obvious that the appearanceof adaptivity requires the existence of a certain levelof complexity, we do not know which type of com-plexity is required for adaptive capacities. In otherwords, how do we draw the boundaries between adap-tive and non-adaptive complex systems? If we are nolonger to believe in rigid boundaries, what makessome systems more adaptive than others? In addition,what types of transition determine the increase incomplexity of adaptive mechanisms? How is norma-tivity expressed through these transitions?

A naturalistic account of adaptivity should addressall these questions from two complementary perspec-tives: (a) an evolutionary perspective, which shouldaccount for the diachronic emergence of adaptivity(what types of evolutionary transition permit the appear-ance of adaptive capacities); (b) an organizational per-spective, which should account for the synchronicemergence of adaptivity from the bottom up (howadaptivity is sustained and enabled by underlying, morefundamental, processes). At the same time, the answershould be grounded on the available scientific knowl-edge and be capable of providing productive feedbackto science both at empirical–analytic and construc-tive–synthetic (biorobotics and simulation of adaptivebehavior) levels.

The structure of this article is as follows. In Sec-tion 2, we introduce the notion of basic autonomoussystems to characterize living beings (i.e., systemscapable of defining and maintaining their own iden-tity). Autonomy provides the naturalized groundingfor normative functionality, teleology, agency, andother fundamental concepts for the understanding ofadaptivity. In Section 3, we focus on the appearanceof the nervous system within basic autonomous sys-tems, the transformations it requires and induces inbiological organization leading to the notion of ahierarchically decoupled dynamic domain in controlof adaptive behavior. We conclude with some funda-mental properties of adaptive behavior that havebeen closely studied in recent adaptive science:embodiment, situatedness, and emergent functional-ity.

2 Autonomy and Intrinsic Teleology

It seems intuitive that adaptivity implies the distinc-tion and the specific relationship between a systemdoing something by itself—an agent—and an environ-ment. This fundamental dichotomy is very often takenfor granted and the characterization of adaptivity isreduced to the establishment of the type of relation-ship between an agent and its environment. However,if we are to proceed from the bottom up, not takingany distinction for granted, we must first ask whattypes of natural process can constitute a system thatproduces its own identity by separating itself as differ-ent from its environment and establishing some func-tional interactions with it.



Barandiaran & Moreno Adaptivity: From Metabolism to Behavior 327

2.1 Self-Maintaining and Far-From-Equilibrium Organization of Minimal Autonomous Systems

A brief examination of the types of process that sur-round us reveals that the universe has evolved by pro-ducing forms of order (such as rocks or solar systems)in some places, whilst in others matter shows noapparent cohesion whatsoever (such is the case ofgases,2 for instance). Ordered matter takes two differentforms: in some cases, basic components appear lumpedtogether constituting conservative static structures, andin others they constitute dissipative dynamic structures.

The first type refers to spatially ordered forms ofassemblage of material subunits, where this order istemporally instantaneous, such as in rocks or crystals,or temporally unfolded, such as atoms or planetarysystems. In both cases, the form exhibited is just anexpression of the intrinsic nature of a set of compo-nents that interact under certain conditions and thatwill exist indefinitely once created (i.e., energeticallyall these systems are conservative).

The other form of order is dissipative, and appearsin far-from-equilibrium (FFE) conditions. This form oforder is found among so-called “dissipative structures”(Nicolis & Prigogine, 1977). A dissipative structure3 isa set of interacting elements that generate a cohesivedynamic pattern in FFE conditions. Examples of thistype of system are whirlpools, hurricanes, oscillatorychemical reactions, and living beings. What all thesedifferent systems have in common is the fact that ahuge amount of microscopic elements adopt a global,macroscopic ordered pattern in the presence of a spe-cific flow of matter and energy (i.e., under certain bound-ary conditions). Interestingly, their internal dynamiccohesion is not only a consequence of the material fea-tures of their components but also of the achievementand maintenance (in FFE conditions) of some type ofcircular causality, as the very macroscopic pattern con-tributes to the maintenance of dynamic cohesion at themicroscopic level. These systems are able to generateand maintain, through recursive dynamics, a type ofcorrelation among their constitutive elements that oth-erwise would remain disconnected. This recursivity isprecisely what provides a minimal form of self-createdidentity.

Now, it is obvious that whirlpools, hurricanes,and oscillatory chemical reactions are too simple toproduce any specific and distinctive form of interac-

tive process with the environment. Thus, in order todiscover the origin of adaptivity, we have to look forforms of (self-)organization capable of evolving andgenerating complex and diverse ways of self-mainte-nance. Among the wide set of self-maintaining organ-izations, those based on chemical processes are ofparticular interest, because they allow the constructionof complex recurrent organizations through the crea-tion of local and selective constraints. This type ofspecific constraint is possible at the chemical levelbecause of the action of the shape and reactive capaci-ties of its constituent components, the molecules.4 Thisis the organizational framework of the early prebioticevolution.

During the process that gave rise to life, a funda-mental step was the appearance of systems whose pro-ductive activity included the construction of a selectiveand functionally active membrane. This change led to aprogressive takeover of (at least part of) those boundaryconditions, thus ensuring the maintenance of the sys-tem. In other words, a form of organization appearedin which the conditions of its maintenance wereactively controlled by the very organization, thus cre-ating a first and self-sustained separation between sys-tem and environment. We call these autonomous(from the Greek auto-nomos, self-law) because (someof) the constraints that define the dynamics of the sys-tems are the result of its very organization. In otherwords, what determines the behavior of the system isnot just the physical laws and a set of externallydefined boundary conditions but the capacity of thesystem to redefine (part of) its boundary conditions(Moreno, Etxeberria, & Umerez, 2008).

Actually, this concept of (minimal) autonomy isvery similar to the idea of autopoiesis developed byMaturana and Varela (1980). Our emphasis, however,focuses on the FFE and thermodynamically opennature of these systems, from which a crucial implica-tion follows: interactive dynamics are constitutive ofthe system and not something to be added, a posteriori,in the form of a structural coupling. For Maturana andVarela, autonomous systems are defined by the abstractproperty of operational closure, leaving aside materialand energetic requirements. As a result, the environ-ment appears only as a source of perturbations due tothe concurrent structural coupling between the systemand its environment. What matters for autopoiesis isthen the conservation of autopoiesis, the compensationenvironmental perturbations or the accommodation of




non-disruptive deformations. For a constitutively opennotion of autonomy, it is through the flow of matterand energy, required for the appearance of FFE dissi-pative structures, that the system achieves its unity.Therefore, by definition, the systems appears thermo-dynamically “hungry,” in need of coupling with theenvironment, which is no longer a mere source ofuncomfortable perturbations to be compensated but thesource of an essential flow. (The consequences forcharacterizing adaptive behavior will be crucial as sys-tem–environment interactions will become function-ally integrated on the very process of becoming anautonomous adaptive system.)

Two constitutive processes can be differentiatedwithin this concept of minimal autonomous organiza-tion5 (see Figure 1a):

1. Constructive processes are those that participatein the continuous production of the system (e.g.,chemical reactions). Constitutive processes arenetworked, thus creating a closed organization wecall the constructive cycle. The constructive cycleis defined by the production of a set of constraintsthat recursively regenerate the conditions of theirproduction. The network of reactions that pro-duces the components of the network itself(metabolism) constitutes the most basic example

of a constructive cycle in a minimal autonomoussystem. However, given that this network onlyexists as a thermodynamically dissipative organiza-tion, its maintenance requires interactive processes.

2. Interactive processes are the processes generatedby the constraining action exerted by the construc-tive cycle on the flow of matter and energy betweenthe system’s boundary and the environment6 so asto ensure the system’s maintenance. At a molec-ular level, interactive processes require systemsendowed with a physical separation (a membrane)between their constitutive organization and theenvironment. An example of a minimal interac-tive process is active transport through the mem-brane, breathing, or adaptive behavior.

Practically all forms of present-day life are capa-ble of performing a wide range of processes in theirenvironments in order to ensure their maintenance.Living organisms at large define their own identity anddifferentiate themselves from their environments. Theway in which they do this is through their metabolicorganization: a self-producing network of chemicalreactions that controls some of its boundary condi-tions. Thus, the basic organization of all present-dayliving beings is essentially what we have elsewherecalled basic autonomy7 (Ruiz-Mirazo & Moreno, 2004).

Figure 1 Conceptual diagrams of increasingly complex organizations. (a) Diagram of constitutive processes in basicautonomous systems; the constructive cycle requires a flow of matter and energy that is actively constrained by the sys-tem, giving rise to interactive processes and thus bringing about a minimal form of agency. (b) Adaptive systems show apartially decoupled mechanism that regulates interactive and constructive processes adjusting and switching betweendifferent alternatives according to external perturbations and conditions. (c) Adaptive agents appear when the interac-tive processes become a cycle. (Copyright © 2007 Xabier Barandiaran under a CreativeCommons Attribution – ShareAlike license: http://creativecommons.org/licenses/by-sa/3.0.)




2.2 Normative Functionality and Intrinsic Teleology Naturalized

Minimal autonomy thus described already provides anaturalized criterion for functionality and normativityin natural systems. The issue of the attribution of nor-mative functionality to natural systems is of majorimportance for the task of naturalizing adaptivity. Asobservers, we can define systems arbitrarily. Forinstance, if we are to use differential equations to builda mathematical model of a system we must (a) choose aset of observables, (b) create some variables represent-ing those observables in our model, and (c) performoperations on the system to establish how variations onan observable produce variations on other observables,so that (d) we can abstract a set of differential equa-tions that specify the rate of change of the variables inthe model. If the model adequately predicts the func-tioning of the system we have an adequate model. Atthis point, we can attribute a “way of functioning” tothe system that corresponds to the functions of ourmodel. Nevertheless, we are faced with the two prob-lems presented in the introduction: (a) how to justifythe selection of observables belonging to the system(attribution of identity); (b) how to justify that the sys-tem not only functions in a certain way but that, inaddition, it “must function” that way and not inanother (attribution of normative functionality; seeMillikan, 1984 for an influential conception of thisproblem). For artificial systems, we, as designers orusers, can attribute a certain goal to the system accord-ing to our intentionality. Under the attributed goal orpurpose, we claim that the machine works properly orthat it is broken and malfunctions, although the goodor bad functioning of the system (the machine) is com-pletely extraneous to the structure of the machineitself. Equally, what belongs to the system and whatshould be left out (as irrelevant) is determined by thegoal we project upon the system; that is, what is an“essential” part of the system (as distinct from its envi-ronment) is defined in relation to the desired function-ality that we as designers or users expect to achieve ina set of contexts.

In contrast, living systems in general and cogni-tive systems in particular are capable of definingthemselves (as we have explained above) and, morespecifically, of determining their own normative func-tionality; that is, what is good or bad (right or wrong)for them does not depend on an external observer,

designer or user but on their own organization. Morespecifically, in autonomous systems, a process (con-structive or interactive) is functional if it contributesto its self-maintenance (Bickhard, 1993; Collier,1999). A process, in turn, becomes normative if it isdynamically presupposed by other processes in theircontribution to the overall self-maintenance of anautonomous system (Christensen & Bickhard, 2002);that is, a behavior or internal component is dynami-cally coupled with the rest of the components so thatthe overall maintenance of the whole organizationdepends on it. Normativity8 refers to the fact that a setof processes that constitute the system must occur asthey do in order for the system itself to exist. A basicexample of normative (proper, necessary) functional-ity is active transportation through the cell membrane.This process becomes normative because the level ofchemical concentrations that the membrane’s activetransport retains within the cell is necessary for somemetabolic reactions to maintain the appropriate rate tosustain the network of reactions, which in turn pro-duces the membrane, and so on in a circular and inter-dependent manner. At a higher level, the normativefunction of the lung is oxygen intake, because thedynamic-metabolic organization of the rest of theorganism relies on this oxygen intake for its function-ing and existence. This type of circularity is character-istic of autonomous systems: a set of networkedcomponent processes that depend recursively on eachother, so that the system, as a whole, is cause andeffect of itself.

Because of this circularity in autonomous systems,identity and normative functionality are not observer-dependent but intrinsically causal: The network as awhole (the very system) will not exist in the absence ormalfunctioning of the component processes (given itsFFE nature and the circular dependency between proc-esses). In other words, in autonomous systems what-the-system-does (the way it functions) and what-the-system-is (its structure) are highly intertwined andthey merge together in its organization.

The holistic, integrated, and self-maintaining organ-ization of autonomous systems has some importantconsequences for the way they are described. Forinstance, the use of teleological terms to characterizetheir functioning can be naturalized; unlike its use todescribe some artifacts that perform a goal-seekingbehavior, such as thermostats or target-seeking mis-siles. These are artifacts that have been designed to




correct their behavior (usually by a negative feedbackmechanism) according to an externally defined goalstate. Expressions such as “the purpose of the thermo-stat is to maintain the room temperature at 23 ºC” areused as metaphorical shortcuts to describe the behav-ior of such systems. However, what the goal state isremains completely extraneous to the mechanism thatachieves it; the system is independent of the goal stateor set of parameters it controls (which are externallyimposed). Autonomous systems are different. Theirexistence depends on the FFE stability they produce.The stability point or set of points that enable a systemto exist are its goal states. These goal states are notjust goal states because the system compensates fordeviations from them, but because the goal state is thecondition of possibility of the system itself. In otherwords, in autonomous systems, the goal state of thesystem and the organization that instantiates it are oneand the same thing. Autonomous systems have animplicit teleology9 as their internal causal circularitymakes each process of the system a contribution to itsglobal self-maintenance.

Thus, the basic type of autonomy just described isthe lower level and most fundamental type of autonomy,that of material and thermodynamic self-constructionand self-maintenance, constitutive of all living beings,upon which higher levels of autonomy appear.10 Basicautonomy generates a cascade of emergent propertiessuch as identity formation, normative functionality,and implicit teleology. By expanding the above analy-sis and properties to the interactive cycle of autono-mous systems, we can naturalize a set of characteristicsthat are necessary to describe adaptive behavior. Theseare agency, adaptivity, and explicit teleology.

3 Adaptivity and Agency

From a dynamic point of view, we can abstract a set ofboundary conditions and an essential parameter valueregion to be necessary for the maintenance of a FFEsystem. Based on the definitions of Ashby (1952), wecall these parameters and boundary conditions “essen-tial variables,” and the range within which the sys-tem’s organization can be maintained “viabilityboundaries.” The FFE nature of autonomous systemsmakes at least one of their essential variables have anintrinsic inertia toward outside the viability bounda-ries. Some of the essential variables are also non-con-

trolled variables, in the sense that no change of theinternal variables of the system can directly control itsstate. As a consequence, only the coupled system–environment can maintain the essential variableswithin viability boundaries (hence the importance ofinteraction processes in autonomous systems).

In general, among the system–environment rela-tionships, and from the point of view of the effect ofthe process on the system, some processes are func-tional for the system when they contribute to maintainessential variables within their boundaries of viability.Some others can be dysfunctional when they “push”essential variables outside the boundaries of viability.Many others are neutral (they have no effect on thestate of the essential variables). However, from thepoint of view of the cause of these processes, we canalso classify them as active (if the processes are trig-gered by the system as a whole) or passive (when theinteraction is induced from outside or the process isthe result of physico-chemical laws without any inter-nally generated constraint acting upon the process;e.g., osmosis). We use the term “agents” for those sys-tems that interact with their environments, so that thechanges produced between the system and its environ-ment contribute to its self-maintenance (see Figure 2for an inclusive classification of the different types ofprocess). The fact that a coupled process can beactive, rather than passive, is of fundamental impor-tance for the task of naturalizing agency, and itdeserves a more precise definition. By active wemean, in negative terms, that the interaction is pro-duced neither by an external source nor by means ofunconstrained physical laws (i.e., spontaneously andindependently of the particular organization of thesystem). In positive terms, an active process is onethat makes use of the organization of an autonomoussystem to produce a constraining effect to ensure itsown self-maintenance. Actions require work, andwork requires suitable energy. Thus, at the minimallevel, active interactions require that the system chan-nels the thermodynamic flow of its FFE organizationinto the creation of specific constraints.11 In addition,another crucial factor that characterizes active proc-esses is the complexity asymmetry of the coupledprocesses, laden to the side of the autonomous system(i.e., the interactive mechanism of the system is,because of its holistic organization, more complexthan the coupled process it sustains). “More complex”means that the set of variables internal to the agent are




more functionally differentiated and integrated (Tononi,Edelman, & Sporns, 1998) than those of the environ-ment involved in the dynamic coupling (for a quanti-tative comparison between internal and behavioralcomplexity, see Seth & Edelman, 2004).

Robustness is the capacity of a system to maintainits organization in the face of internal and external per-turbations. This capacity can be achieved by a numberof increasingly complex mechanisms: a simple buffer-ing mechanism, a distributed structural stability or anadaptive regulatory subsystem. In any case, a sense ofmargin of viability is required; otherwise even thesmallest perturbations will destroy the system’sorganization. The simplest forms of self-maintainingsystems, such as hurricanes or candles, cannot really“do” anything in order to compensate for new envi-ronmental conditions, they only have a margin ofmaintenance provided by a buffering or flexible struc-ture. However, more complex forms of self-maintainingsystems, such as hypothetical cellular protometabo-lisms, achieve robustness only by using their internalorganization to constrain some of their boundary con-ditions. In other words, to sustain self-maintenanceunder different environmental conditions, they must“recruit” their internal organization to regulate thecoupled processes, which now become interactiveprocesses. Interactions are thus active and functionalprocesses between the system and its environment.Active ion pumping by the membrane could be anexample (Moreno & Barandiaran, 2005; Moreno &Etxeberria, 2005; Ruiz-Mirazo & Moreno, 2004). Thisinteraction becomes functional because, as a result ofion-pumping, it produces an internal change that isnecessary for the ongoing activity of the constructivecycle (reduction of ion concentration in the interior ofthe cell).

The most simple forms of autonomy are structur-ally stable, in the sense that they can compensate inter-nal and external perturbations by means of a self-regulating mechanism that is integrated and distrib-uted over their constitutive organization; that is, theregulatory mechanisms are embedded in constructiveand interactive processes and so regulatory and regu-lated processes are indistinguishable. The property ofstructural stability12 corresponds with the idea of con-servation of autopoiesis as originally formulated byMaturana and Varela (1980). However, when FFEsystems increase their complexity13 they become morefragile; noise and environmental perturbations affect

Figure 2 Inclusive classification of different processes inautonomous systems. Functional processes are thosethat contribute to self-maintenance, and these can beboth internal to the organism or interactive with the envi-ronment. Active processes are those generated by con-straints produced by the system. Within active processes,interactive processes give rise to agency; these are theinteractive processes that are functional in virtue of thechanges induced in the environment. Adaptive processesare those active and functional processes that are carriedout by a specific mechanism, which detects and compen-sates tendencies of essential variables when they are suf-ficiently close to the boundaries of viability. Adaptiveprocesses can be internal or coupled. Interactive andadaptive processes give rise to adaptive agency. Finally,behavior appears as a subclass of adaptive agencybased on mechanical articulation. (Copyright © 2007 Xa-bier Barandiaran under a CreativeCommons Attribution –Share Alike license: http://creativecommons.org/licenses/by-sa/3.0.)




their organization, which, given its holistic nature, dis-integrates easily. Thus, as complexity increases evenmore, self-maintenance under relatively wider condi-tions requires special mechanisms capable of ensuringadequate coupled processes. Structural stability cannotensure the maintenance of complex systems under dif-ferent, variable environmental conditions. Hence, inearly living systems self-maintenance requires an activecontrol of interactive and constructive processes bymeasuring different conditions and monitoring its ownconstitutive processes so as to avoid or prevent dys-functional situations. In other words, in addition toexecuting a constructive or interactive process thatcontributes to self-maintenance, the system is alsocapable of switching between different alternatives,adjusting them, and so forth, according to externalchanges. This is the essence of adaptivity recentlydefined by Di Paolo (2005):

A system’s capacity, in some circumstances, to regulate

its states and its relation to the environment with the result

that, if the states are sufficiently close to the boundary of

viability, a) tendencies are distinguished and acted upon

depending on whether the states will approach or recede

from the boundary and, as a consequence, b) tendencies of

the first kind are moved closer to or transformed into tenden-

cies of the second and so future states are prevented from

reaching the boundary with an outward velocity. (p. 438)

Adaptivity requires the establishment of an explicitnormative regulation. As Di Paolo maintains, in non-adaptive self-maintaining systems the natural distinc-tion between self-maintenance and disintegration isnot yet accessible to the system, unless it is also ableto regulate itself with respect to a norm. Whereas inpre-adaptive systems self-maintenance depends on therange of values that the essential variables can with-stand, adaptive systems have the capacity to modulate(internally and interactively) the trajectories of theessential variables of the constitutive processes (unlikestructurally stable systems; see Figure 1b). Essentially,adaptivity requires a regulatory control over the basicfunctioning of the system. Therefore, in order to beadaptive, a system must be organized such that thereis a relative decoupling between the dynamics of a reg-ulatory subsystem and that of its basic constitutiveorganization. This has important consequences, as someof the properties traditionally assigned to autopoieticsystems cannot be derived from autopoiesis alone but

presuppose adaptive capacities (as argued in detail byDi Paolo, 2005). In addition, in most cases, adaptiveregulation takes place not just by transforming out-ward tendencies into inward trajectories (i.e., not justby avoiding negative tendencies) but by actively seek-ing to improve the state of essential variables so thatregulation takes place not just in reference to theboundary of viability but graded and directed by a“sense of well-being.”

Adaptivity is a capacity that all present-day livingorganisms possess.14 The simplest mechanisms ofadaptive regulation fall into two different categories.One is exemplified in the Operon activation and deac-tivation of genes as a switch between metabolic path-ways according to certain environmental conditions.The other is constituted by a whole subsystem of bio-chemical pathways not directly involved in the basicself-constructing metabolic network (as is the case ofchemotactic agency in Escherichia coli; see examplesat the end of this section). However, the commoncharacteristic of both cases is that some degree ofdynamic decoupling from the basic constitutive proc-esses is required. In the first case, metabolically off-line gene-strings act as instructive switches betweendifferent metabolic pathways. In the second case,chemical pathways that are independent of the basicmetabolic-constructive cycle sustain the interactiveloop. This decoupling of genetic regulatory mecha-nisms from the basic metabolic network allows a selec-tive choice among a large amount of not yet functionaldynamical states of the constitutive self-maintainingmetabolic network. These decoupled systems open thepossibility to consistently speak in terms of an inter-nally generated mechanism for normative regulation.The capacity to differentiate between, and compensatefor, tendencies requires that whatever makes a distinc-tion and generates a compensation be dynamically dif-ferentiated from what it distinguishes and acts upon(which presupposes operational mechanisms to distin-guish between the different implications of equallyviable paths of encounters with the environment).

Thus, we are talking about two dynamic “levels”in the system: the constitutive level, which ensuresongoing self-construction, and the (now decoupled)interactive subsystem, which regulates boundary con-ditions of the former (e.g., concentration gradient ofmetabolites across the membrane through displace-ment to richer environments). It is clear that decou-pling permits a specific phenomenon: independence




of low-level functionality (constructive processes)from high-level variation (state of regulatory mecha-nism). Thus, free from the lower-level constraints,higher level variation can be left to “spontaneous”dynamics, provided that a further coupling is estab-lished linking higher-level to lower-level states in afunctional way so that this higher-level variation (orfreedom) is in turn functionally recruited to servelower-level functions. This way, certain states of thelow-level network will be stabilized when contribut-ing to the creation of new forms of self-maintenance.

We can naturalize the claim that some interactionor process is perceived as bad or good by, and for, thevery system (and not only by, and for, the externalobserver); that is, this good or bad functioning for thesystem is objective because it is detected and compen-sated by the system, in an effective, functionally inte-grated way. Thus, adaptive systems are an instance ofexplicit teleology as, in addition to having an intrinsicgoal (because of their basic autonomous organiza-tion), they also act according to this goal, generatingglobal constraints, over their minimal basic organiza-tion, so that a meta-regulatory process emerges.

Adaptivity takes two basic forms depending onwhether the mechanisms of regulation take place atthe constructive or interactive level. In the first case,internal or external perturbations are compensated byadjusting or transforming constructive processes (suchis the case of the Lac operon mechanism). The secondform of adaptivity turns out to be of particular interestbecause it gives rise to adaptive agency: adaptation toperturbations is achieved through recursive interac-tions with the environment so that interactive proc-esses become a cycle (see Figure 1c). Interactionsbecome functional in virtue of the changes inducedoutside the system or, more specifically, in the rela-tionship between the system and its environment.Motility-based interactions are the most clear exam-ples of this type of functional feedback through theenvironment. A paradigmatic case is bacterial chemo-taxis: in order to change its current environment toanother that contains more nutrients, bacteria engageon a sensorimotor cycle. The set of interactive proc-esses performed by the bacteria become cyclicbecause it is realized by modulating effector processesaccording to the detected conditions of the environ-ment in a recursive manner. However, there are alsoother forms of adaptive agency not based on motility.A plant, for instance, can become an adaptive agent by

secreting and detecting chemical substances, engagingon a defense–attack coupling with certain fungi on itsproximal environment. In these cases, the transforma-tions induced on the system–environment relationshipalso become functional inputs for achieving adaptiveregulation; hence the notion of “interactive cycle” as adistinctive character of adaptive agency (unlike casesof internal adaptive processes or single functionalactions into the environment, such as ion-pumping).

The appearance of adaptive agency implies theemergence of detection and effector mechanisms bywhich the adaptive regulation is linked with the envi-ronment and the constitutive organization of the auton-omous systems. In addition, two regulatory processesappear linked to each other: the adaptive regulation ofthe essential variables through recursive interactionswith the environment on the one hand, and the regula-tion of this interactive cycle according to its effects onthe essential variables on the other. This intertwinedregulation along with the detection–response couplingwith the environment brings forth a genuine type of tele-ological agency, in which explicit teleology is extendedinto environmental interactions.

We can now introduce the concept of motility asthe capacity of an agent to move by its own means, sothat it is able to execute fast (relative to its size) direc-tional movements to change the environment lookingfor preferred conditions. In the case of adaptive motil-ity, detection of and functional response to environ-ment-relevant changes becomes a “sensorimotor”cycle, the viability of which is strongly affected bysize–time constraints. It is this high size–time (speed)constraint that characterizes sensorimotor adaptabilityfrom other forms of adaptability.

Let us summarize the proposed distinctions byproviding some characteristic examples.

1. Simplified bacteria. Buchnera aphidicola (Moya,Peretó, Gil, & Latorre, 2008; Shigenobu, Watan-abe, Hattori, Sakaki, & Ishikawa, 2000) can beconsidered to be a living system in its own rightas the machinery of self-(re)production and ener-getic and material exchange with the environmentis at work, but should be considered to be neitheran adaptive nor an interactive agent. This bacte-rium is an extremely degenerated organism. Itlives as an endosymbiont inside the bacteriome ofcertain insects where the local environment ishighly homogeneous and co-adapted to its meta-




bolic organization. It obtains the necessary ener-getic and chemical input for self-maintenance fromits medium with almost no change of environmentalconditions. As a consequence, adaptive mecha-nisms in Buchnera aphidicola are highly degener-ated and these organisms are considered to be closeto a structurally stable minimal-autonomous sys-tem (to the extent that they are currently used asmodels organisms for the study and synthesis ofminimal life forms).

2. Lac operon mechanism in E. coli. The normalmetabolism of E. coli depends on the presence ofglucose in its environment. However, when thelevels of glucose in its environment become verylow and another sugar (lactose) is abundant in theenvironment, a mechanism called lac-operon isactivated: The detection of lactose triggers theexpression of certain dormant genes,15 which inturn instruct a new metabolic pathway that metab-olizes lactose. This metabolic mechanism is adap-tive because it implies a (meta)regulation of theinternal constructive processes according to thedetection of a certain environmental condition (thepresence of lactose and the absence of glucose),which jeopardizes the self-maintenance of the sys-tem. Nonetheless, this form of adaptivity does notimply agency as the changes produced for self-maintenance are mainly internal.

3. Chemotaxis in E. coli. E. coli alternates tumbling(random rotation) and forward displacements,generating a net displacement toward environ-ments with increasing concentrations of metabo-lites. This is achieved through a “two componentsignal transduction” mechanism, capable of meas-uring the temporal difference of attractant concen-trations in the environment and changing thefrequency of the flagellar rotation accordingly (seevan Duijn, Keijzer, & Franken, 2006 for a moredetailed analysis). In this form of adaptivity, thesystem interacts functionally with its environment(moves up-gradient until the necessary level ofsugar is encountered). The interaction is functionalin virtue of the transformations induced in theenvironment (the concentration of sugar increasesin relation to the system) and the system operatesrecursively on these environmental changes (sugar-detection and frequency of flagellar rotation arecorrelated). Among an open set of environmentalvariables, the sugar concentration gradient shapes a

functional world of interactions for the bacteria’sself-maintaining capacity.

In short, some fundamental elements, such as func-tionality, normativity, and teleology, are the conse-quence of the autonomous organization present in thesimplest biological systems. Adaptive agents (thoseautonomous systems capable of interacting with theirenvironments by detecting and compensating for ten-dencies of their essential variables) extend an explicitteleology to their environments. As seen in the nextsection, biological organizations that support this typeof agency through biochemical mechanisms are severelylimited in their capacity for open-ended agential com-plexity. Thus, without a qualitatively different internalorganization of autonomous systems, adaptive agencyeventually exhausts its capacity for complexity growth.The appearance of the nervous system (NS) solvesthis problem, leading to a whole set of bodily changesand to a qualitatively different organization of adap-tive agency.

4 Nervous System and Behavioral Agency

4.1 Limitations of Agential Mechanisms at the Metabolic Level

As we have seen, the appearance of adaptivity impliesa certain degree of decoupling within the organizationof the agent; there should be mechanisms that supportthe interactive processes whose functioning is rela-tively independent of those involved in the construc-tive processes. This internal organization of adaptiveagency therefore requires a relatively complex metab-olism, instructed and regulated in somatic time by anoff-line conservative structure (DNA) or by a differentand relatively independent subsystem of chemicalreactions (van Duijn et al., 2006). These early adap-tive mechanisms are context-specific regulatory mech-anisms and mostly genetically specified.

However, the use of biochemical mechanismssupporting interactive tasks severely limits the capac-ity to achieve increasingly complex forms of agency.The reason is that there is a serious bottleneck in theevolution of movement-based agency supported by bio-chemical mechanisms. First, the bottleneck appearsbecause the level of complexity that the adaptive sub-system can achieve (within the biochemical medium)




without severe interference with metabolic processesis very limited. Second, as the size of the organismincreases, the fast and plastic correlation between sen-sor and effector surfaces becomes harder (or evenimpossible in multicellular organisms) as a result ofthe slow velocity of diffusion processes. Third, thereis also the problem of achieving unified body coordi-nation for displacement. The type of rather sophisticatedmotile agency displayed by Paramecium illustrates thetension produced by the combination of these three fac-tors: epithelial conduction (through Ca channels) isused to solve fast and coordinated beating of ciliabecause, unlike the case of E. coli, it could not beachieved by mere diffusive mechanisms. However,the complexity (in terms of functional diversity andintegration) that homogeneously spreading epithelialconduction can achieve is very limited.

The appearance of multicellularity posed an impor-tant challenge in the evolution of agency, as at this sizebiochemical mechanisms cannot support fast and ver-satile motility. There are two causes of this problem:the enlarged internal distance between parts of thebody, which needs to be connected in short delays (sothat the organism can move fast and coordinately),and the need to modulate the organization of connec-tions selectively (to obtain the adequate sensorimotorcorrelations) for versatile, plastic agency. Hence, ifbiochemical network plasticity were the only mecha-nism for accomplishing adaptive interaction and self-maintenance, the forms of movement-based agencywould probably be very limited at the multicellularsize.16 However, when in the development of somemetazoans a new type of cell (the neuron) started todifferentiate itself, this limitation could be overcome.Neurons differentiate as cells capable of formingbranches, interconnected through plastic electrochem-ical pathways and capable of propagating and modu-lating electric potential variability.17 In fact, theseinterconnected cells led to the establishment (about600 million years ago) of a dynamic network capableof managing an efficient coordination between sensorand motor/effector structures in multicellular organ-isms (Llinás, 2001).

4.2 Hierarchical Decoupling of the Nervous System

Since the very beginning of its evolution, neuralorganization appeared as an extended network capable

of producing a recurrent dynamic of specific patternsindependent of the underlying metabolic transforma-tions that the organism undergoes. Unlike chemicalsignals circulating within the body, which directlyinteract with metabolic processes because of their dif-fusive nature, the electrochemical interactions betweenneurons make open-ended recurrent interactions withinthe NS itself possible. The NS constitutes a cellularinfrastructure that converts metabolic energy intofinely modulable electrodynamic processes, thus cre-ating a new dynamic level almost free from the thermo-dynamic constraints characteristic of the biochemicallevel of metabolic-constructive processes. What makesneural interconnections so special is that they create anincredibly rich and plastic internal world of patterns offast connections, hierarchically decoupled from themetabolic processes.

As we have pointed out elsewhere (Barandiaran,2004; Barandiaran & Moreno, 2006; Moreno & Etxe-berria, 2005; Moreno & Lasa, 2003), the hierarchicaldecoupling of the NS from metabolism means thatmetabolism generates and sustains a dynamical system(the NS) minimizing its local interference with it. Theterm “hierarchical” refers to the fact that metabolismproduces and maintains the architecture of the NS byproviding the necessary energy to feed its dynamics.However, the term “decoupling” means both (a) thatneurons minimize interference in their local metabolicprocesses with their ion-channeling capacities and (b)that the metabolic-constructive organization of theorganism (digestion, circulation, etc.) under-deter-mines the activity of the NS, which depends on itsinternal dynamics and its embodied sensorimotor cou-pling with the environment. Operationally speaking, ifwe are to predict the state of the NS, hierarchical decou-pling means that neither local states of cell metabolismnor the state of metabolic organs alone are going to bevery useful; on the contrary, the electrochemical statesof other neurons and their embodied sensorimotor cou-pling with the environment might provide a much bettermodel for prediction. In other words, the biophysicalspecificity, high connectivity, embodiment, and situat-edness of neural electrochemical dynamics make itirreducible to the metabolic substrate of its constituentcomponents (the neurons) and the organismic proc-esses of self-construction and repair (state of otherbody organs and processes involved).18

Hierarchical decoupling permits the specificationof a set of operational primitives (the lowest level




dynamical observables necessary and sufficient tomodel nervous activity) and their relationships, whichconstitute the control mechanisms of animal agency.It is commonly accepted that the primary operationalprimitives are the change of membrane action poten-tials of neurons over time (that generally take the formof spikes), which conserve dynamic variability interms of spike frequencies and time distance betweenspikes. Synaptic connections, however, specify a con-nectivity matrix (the transformation functions betweenprimary operational primitives) while neural modula-tors (local and global synaptic modulators and actionpotential threshold modulators) become secondaryoperational primitives (as they become operationalprimitives in virtue of their effect on the spikes). Thesearch for these dynamic primitives and their func-tional higher level causal organization constitutes thesearch for a neural “code:” the set of primitive variablesand relationships that constitutes the dynamic domaincapable of modeling adaptive behavior. In other words,the type of local operational differences that can make asystematic global difference in behavior (spike rates,interespike intervals, time of arrival, gas-net modula-tion, synaptic modulator, axonal growth, etc.).

The decoupling of neural processes from theunderlying metabolic processes raises the question ofhow to characterize this specific dynamic domain. Aswe have seen, the active electrochemical conductivityof the components of the NS (the neurons) is organ-ized in spikes or action potentials, thus generating anew dynamic domain that is built up tangentially tothe metabolic processes of the organism, althoughrealized and sustained by it. At the same time, the all-or-nothing characteristic of neural spikes allows for astable combination of these, which, added to the net-work structure of the NS and the action of neuralmodulators, generates a high-dimensional, non-linear,recurrent, and recursive domain. Nonlinearity allowsdistinctiveness of states, while recurrency, provided bythe structure of the network, allows circularity or reentry(Edelman, 1987). Recursivity, however, takes placebecause spikes can affect themselves through the neuralmodulators they activate. As a result, the effectivedimensionality of the system is constantly being rede-fined by its own activity. Thus, unlike the aforemen-tioned forms of decoupling within the biochemicalorganization, which allow primitive forms of adaptiveagency, the hierarchical decoupling of the NS permitsan open-ended growth of complexity in the forms of

agency (impossible to achieve by previous agentialmechanisms).

In addition, there is a causal link between the neu-ral domain and certain external processes that belongto other dynamical levels (e.g., metabolic processes inthe muscles). This causal connection is largely inde-pendent of energetic or material aspects, as neuralstates produce changes in body states by formal ratherthan energetic means in a type of lock-and-key causal-ity (what we call formal causality). This causal link isestablished through the pattern of spikes and notthrough the energetically determined causality throughwhich these very patterns of spikes propagate. Thus, forexample, the motor action caused by neural spikes isnot determined by the electrochemical energy thatconstitutes action potentials but by their form or pat-tern, which muscle cells “interpret” (i.e., the processby which the neurotransmitters that neurons generateact by selecting metabolic energy to produce move-ment). This process is similar to the electric patternsthat travel along wires and connect two computers;these patterns produce changes in the terminal not byvirtue of the electric energy they convey but by virtueof the sequence of changes in amplitude and frequency.In other words, the neurotransmitters that neurons gen-erate (when a given pattern of spikes from further neu-rons arrives) trigger a cascade of chemical processesin the muscles that convert patterns of spikes intomechanical work.

Thus, the NS, hierarchically decoupled and endo-wed with the capacity for formal causation, justifiesthe characterization of the neural domain as properlyinformational.19 Neural primitives can be consideredas non-dynamical elements in relation to the underly-ing metabolic processes in the sense that, from thepoint of view of the modeling of metabolic-constructivedynamics of the organism, the NS appears decoupled.From the point of view of musculoskeletal dynamics,the NS acts as a formal control system, independentfrom the particular energetic details of how movementis achieved in the muscles.

4.3 Behavioral Agency

Interestingly, the appearance of multicellular organismsendowed with a (sub)system allowing fast, efficient,and plastic agency, was necessarily accompanied byother important changes in their internal structure. Theappearance and evolution of the NS brought along with




it changes in the organization of internal circulation,the system of fixation, and even the body shape. Thus,the unfolding of multicellular organisms whose wayof life is based on motility requires radically differentinternal organization, namely a new bodyplan thatallows for a whole set of transformations at both theinternal and interactive levels (Moreno & Lasa, 2003).First, as we have already mentioned, there is theappearance of a decoupled non-biochemically basedadaptive subsystem: the NS. Second, there are specifictissues and body structures capable of channeling met-abolic energy into efficient mechanical energy (mus-cles, skeleton, etc.). Fast movement, in multicellularorganisms, is only possible through specialized organs,which directly convert metabolic energy into mechani-cal energy independently of the continuous process ofmetabolic self-maintenance and morphological trans-formations that the organism undergoes by means ofcell growth and reproduction. We denominate the newform of adaptive agency based on motility of thosemulticellulars endowed with a NS controlling a mechan-ical body as properly behavioral agency.20 Thus, behav-ior is actually fast adaptive motility decoupled frommorphological and, in general, metabolic-constructiveprocesses.21

It is precisely the hierarchical decoupling of theNS and its sensorimotor coupling with the environ-ment that makes it possible to study adaptive behaviorin terms of sensorimotor dynamics (as is the case inseveral fields, such as robotics, cognitive neuroscience,or embodied psychology) and qualifies behavior as aspecific phenomenon distinct from generic biology. Incontrast, the explanation of the interactions of plantswith their environments would require the introduc-tion of additional (non-sensorimotor constraints) suchas the rate of growth through cell replication accord-ing to exposure to light, availability of water in theimmediate surroundings and a host of similar agency–metabolism interdependences. None the less, hierar-chical decoupling from metabolism and sensorimotorcoupling with the environment does not mean that themetabolic substrate of behaving organisms is irrele-vant. In contrast to the local decoupling from meta-bolic processes, a global coupling follows, so thatadaptive behavior will ultimately have to satisfy thedemands of metabolism.

4.4 Neurodynamic Constraints, Self-Organization and Adaptive Behavior

The function of the neural domain in the overallorganization of the organism is to achieve behavioraladaptivity (i.e., adaptive maintenance of essential var-iables under viability boundaries through the neuralsensorimotor control of the interactive coupling withthe environment; Barandiaran, 2004). However, thefact that the global dynamics of NS–body–environmentproduce an adaptive maintenance of essential variablesunder viability boundaries does not specify how thisfunctionality is achieved (i.e., what the dynamic organi-zation of the NS is like and how it is related to behavio-ral adaptivity). Because it is impossible for metabolicneeds alone to instruct functionally such a high-dimensional space, the understanding of behavioralagency requires us to explicitly define what types ofconstraint act on the NS generating functional order.

Dynamically speaking, the activity of the NS isdefined by internal and external constraints. Externalconstraints are those that are not the result of theactivity of the NS itself. We can distinguish two maintypes of external or innate22 constraints on the NS.Architectural constraints look like a first candidate forthe “instruction” and generation of order in the NS,and can, in fact, generate highly constrained structuresand dynamics in some primitive NSs.23 These con-straints are the result of genetically triggered anatomi-cal-developmental processes selected in relation to theadaptive sensorimotor interactions they produced whencoupled to body–environment systems. Body signals,however, act as strong perturbations of neural dynamicsthrough specific signals (pain, hunger, pleasure, etc.)that operate by modulating the overall activity of theNS (generally through specific neural modulator path-ways such as the dopaminergic system). The role ofbody signals is to regulate the behavior sustained byneural activity in relation to the adaptive needs of theorganism. These signals shape neural dynamicstoward the satisfaction of certain goals; they are gen-erally originated on metabolic needs and can be con-sidered innate from the point of view of the earlyactivity of the NS.

So far, then, neural dynamics can be capturedthrough the specification of (a) neural primitives, (b) aset of innate (architectural and body signal) con-straints, and (c) sensorimotor interactions. However,in most of the known neural systems the complexity




of the possible neural dynamics appears underdeter-mined by these constraints. Nonetheless, we still findfunctional order. Therefore, different (non-external)principles of order are required to explain them. Morespecifically, the dynamics of the NS enter a process oflocal and interactive self-organization through therecursive activity of neural dynamics and sensorimo-tor interactions. The hierarchical decoupling achievedthrough the electrochemical functioning of neuralinteractions and their capacity to establish a highlyconnected and non-linear network of interactions pro-vides a dynamic domain with open-ended potentiali-ties, not limited by the possibility of interference withbasic metabolic processes (unlike diffusion processesin unicellular systems and plants). It is precisely theopen-ended capacity of this high-dimensional domainthat opens the door to spatial and temporal self-organ-ization in neural dynamics and generates an extremelyrich dynamic domain mediating the interactive cycle,overcoming some limitations of previous sensorimo-tor control systems.

A paradigmatic example of self-organizing pat-terns in the NS is given by central pattern generators(CPGs) where the interaction between neurons on alocal circuit generates robust oscillatory patterns.However, self-organization also appears at the level ofthe coupling between NS, body, and environment. Infact, the effect of sensory perturbations propagatesrecurrently through the network generating muscle con-tractions, which in turn feeds back to sensory neuronsboth through the changes that movement induces inthe immediate sensory environment and through prop-rioceptive feedback. The recurrent embodied couplingof the NS to the environment results in adaptivebehavioral patterns whose functional stability is theresult of the dynamic integration of neural, body, andenvironmental features. Cases of interactive self-organ-ization have been reproduced over more than two dec-ades of robotic research24 and span across almost all thedomains of living behavior (except for a few rigid andcue-bound reflex-like reactive responses).

So let us recapitulate. In large (multicellular) liv-ing systems, sensorimotor agency requires a dynamicdomain fully decoupled from local metabolic con-straints, namely, the NS. This system, embedded in asensorimotor architecture, is organized in terms ofinternal and interactive self-organized processes con-strained by innate architectures and body signals. Thefunctionality of the system is defined by the mainte-

nance of the system within viability boundaries, whichgives rise to the unfolding of adaptive behavior.

5 Situated and Embodied Nature of the Sensorimotor World

At this level of sensorimotor adaptive behavior, wecan fully identify some more fundamental propertiesof natural behavioral processes that are found to be atthe basis of all cognitive processes. Some of these prop-erties might already appear in earlier (even unicellular)agents, especially those based on motility and endowedwith sensory and motor mechanisms (van Duijn et al.,2006), but their full significance shows up in neurallyguided behaving systems.

Behaving systems are situated systems, and theirrelation with the environment is relative to their situa-tion in it in a non-trivial manner. The behaving organ-ism is not coupled to the environment as a WattGovernor might be coupled to a water flow. Sensoryinput is not only a function of the environment and thetransformations that the agent induces in it but a func-tion of the controlled relative position of the agent inits environment (spatially and geometrically struc-tured through motility). This is a fundamental prop-erty that has been ignored in most of the literature onartificial intelligence. For instance, when functionalbehavior is taken to be the result of extracting statisti-cal properties or patterns from a string of predefinedinputs, the consequences are non-trivial. For example,a non-situated system that is reactive (i.e., whose out-put is determined by the instantaneous input and a his-torically non-modifiable internal structure) cannotsolve a non-Markovian task (i.e., it cannot success-fully classify an environmental condition if detectingthis condition requires it to extract a sequential order).In contrast, a situated system with a reactive controllercan transform non-Markovian tasks into Markoviantasks just by means of exploiting its relative positionin its environment (Izquierdo-Torres & Di Paolo,2005).

In addition, functional behavior is defined by thebody in two different dimensions, which we call sen-sorimotor and biological embodiment, respectively(for a detailed discussion, see Ziemke, 2003). Adap-tive robotics has devoted more attention to sensorimotorembodiment, which is a function of bodily properties inrelation to the environmental sensorimotor coupling




of the agent. The body’s physical and mechanicalproperties shape possible interactions and relative posi-tions through enabling biomechanical constraints. Thespace of motor outputs to be instructed by the organismis not a uniform multidimensional space defined by anumber of degrees of freedom. In contrast, motorembodiment defines a biased “landscape” within thatspace determined by the shape, elasticity of joints, rela-tive orientation, and a host of similar body constraints.An extreme case of motor embodiment is given bydynamic walking models (McGeer, 1990), in which,even in the absence of neural control, a mechanical sys-tem (a pair of legs) determines a well-structured envi-ronmental coupling with its environmental surfacegiving rise to coherent and robust walking behavior.Instead, in the same situation, a disembodied approachwould have required an exhaustive control motor out-put anticipating trajectories and a host of feedbackcontrol mechanisms. However, embodied sensory sur-faces define a range of sensory inputs and particulartransformations and the filtering of them. Finally, bothembodied sensory and motor embodied surfaces appearhighly intertwined because of the circular and recursivenature of sensorimotor interactions that have evolvedtogether. We can call these enabling constraintsbecause they bias the potential dimensionality of thesensorimotor coupling so as to enable or facilitateself-organized developmental and adaptive interactions(cross-modal sensorimotor spaces, developmental scaf-folding by bodily changes, structural adaptation withcertain object size and shape, etc.). In addition, andfrom a computational perspective, embodiment alsomeans that much of the cognitive processing is carriedout as embedded in the structure and mechanical func-tioning of sensorimotor processes.

A less commonly emphasized type of embodi-ment is biological embodiment, which defines the eco-logical network of interactive necessities of the agent inorder to satisfy its basic biological conditions of possi-bility.25 Biological embodiment is in continuous feed-back with the sensorimotor flow. In fact, the world ofa behaving organism is not so much an independent,physical world but the coupling of this external worldwith the “internal” world: the dynamics of the con-structive cycle “expressed” through body signals har-nessing neural dynamics so as to satisfy metabolicneeds. The primary function of sensorimotor dynam-ics is therefore to maintain the essential variablesunder viability boundaries. So the primary sensorimo-

tor correlations in the organization of behavior aredefined by the effect of the sensorimotor coupling onthe dynamics of its biological embodiment. In fact, theworld that comes about through biological embodi-ment can be understood as a mapping between thesensorimotor environment and the basic autonomousviability conditions. This is, properly speaking, theadaptive environment that a natural agent defines.

Artificial agents built according to the so-called“autonomous situated robotics principles” show emer-gent behavior arising from real (or realistically simu-lated) perception–action cycles; they are (or tend tobe) able to measure the relevant parameters of theenvironment to control certain degrees of freedom ofthe system from the very situation in which the systemfinds itself (and not from the point of view of an exter-nal observer), and to physically act in it. However,what is lacking in most (if not all) allegedly adaptiveor autonomous artificial systems is, as Di Paolo(2003) has pointed out, a self-concern of the processesthey undergo “because the desired goal is notdesigned by the robot but by the designer.” In otherwords, these robots lack their own normativity, whichis a consequence of their lack of biological embodi-ment. This means that artificially created adaptiveagents are not fully embodied. While artificial embod-iment does not preclude a useful study of adaptivebehavior, it severely limits the attribution of genuineagential or adaptive capacities to these systems. Analternative research program would require us tomodel autonomy at the level of behavior itself, syn-thesizing agents capable of both maintaining theirown behavioral organization through interactions andgenerating a new level of identity and self-mainte-nance at the sensorimotor level, in analogy with basicor metabolic autonomy (Barandiaran & Moreno,2006; Di Paolo, 2003).

6 Conclusions

We have argued that adaptivity requires autonomy.Artificial adaptive systems are, in fact, possible onlybecause normative criteria are externally imposed byhuman beings, who, as living organisms, are autono-mous systems. Autonomous systems create their ownidentity and differentiate themselves from the envi-ronment. However, autonomy is a necessary, but not asufficient, requirement for adaptivity. As we have seen,




the appearance of adaptivity also requires a regulatorysubsystem within an autonomous system, partiallydecoupled from the dynamical organization of theconstitutive processes of the system.

The most interesting aspect of motility is theappearance of new types of interactive process notdirectly involved in the maintenance of the constitu-tive (metabolic) organization of the system. Thesenew interactive processes give rise to the sensorimotordomain. Therefore, sensorimotor processes are neithermere interactive processes performed by an autono-mous system, nor just adaptive agency, but specific,metabolically quasi-independent interactive dynam-ics of movement-based adaptivity.

However, biochemically supported adaptive motil-ity becomes a serious organizational problem as sizeand complexity increases. Thus, full-fledged behavioronly appears when adaptive motility is supported by ametabolically decoupled regulatory subsystem (theNS) embodied on a musculoskeletal architecture, allow-ing fast and plastic functional sensorimotor coordina-tion at the multicellular scale without interference withother constructive processes (metabolism, growth, rep-lication, etc.). Adaptive behavior is therefore bodymovement through neutrally controlled sensorimotorinteractions that satisfies biological constraints:behavior oriented towards survival and reproduction.Although self-organizing patterns appear in the neuraldomain, as a result of internal and external recurrentinteractions, the biological embodiment still defines thenormative framework and teleology of adaptive behav-ior. The organism as a whole is an autonomous agentbut, from the point of view of sensorimotor dynamics,viability boundaries are externally defined (by biologi-cal or basic autonomous needs). In this sense, adaptivebehavior lies within the biological domain.

Finally, the creation, through the NS, of a highlyrich sensorimotor domain decoupled from the func-tioning of the metabolic organization not only allowsan open growth of the complexity of adaptive proc-esses, but also has a collateral consequence: the pro-gressive takeover of both the organization of behaviorand of the body itself by the neural system, openingthe way to the appearance of the cognitive domain.The study of how this new domain has appeared, itscapacity to provide non-metabolic normativity andteleology, and its relation with the biological domainhas been explored elsewhere (Barandiaran, 2007;Barandiaran & Moreno, 2006).

Acknowledgments

Funding for this work was provided by grants 9/UPV 00003.230–13707/2001 from the University of the Basque Country, andBMC2000–0764 and HUM2005-02449/FISO from the Ministryof Science and Technology. In addition, X. Barandiaran has thesupport of the doctoral fellowship BFI03371-AE from theBasque Government.

Notes

1 The search for a minimal organization is meant to be amethodological strategy to find the “essence” of a processthat appears vastly complex in nature. In physics, the Bohrmodel of the hydrogen atom may be taken as a paradig-matic example. The search for minimalism does not ruleout other approaches (e.g., the modeling of a componentsubsystem, such as the visual system of a cat in cognitionor the Kreb cycle in metabolism). What a minimalistapproach permits is a model template to which otherpieces of modeling work can be attached or integrated.

2 Needless to say, at very large scales gases might achieveforms of cohesion on the basis of gravitational forces suchas stellar gas clouds. For a smaller case, however, gasesare states of matter characterized by their lack of cohesiveunity outside containers, or contingent forms that they canacquire for very short periods of time (such as clouds).This is not to say that a gas cannot take any cohesive formunder any condition. As we see, dissipative structures canappear, structuring non-cohesive material substrates undervery specific conditions.

3 A dissipative structure is in fact a self-organizing system.By self-organization we mean that local non-linear inter-actions between components generate a global behavior,which is maintained through a certain number of con-straints of which at least one is a product of the global pat-tern (Ruiz-Mirazo, 2001). Note that the global pattern isnot instructed (dynamically specified) from outside, norcan it be reduced to or predicted from the activity of anyof its local components.

4 Most “purely physical” self-organized systems (such aslasers, spin glasses or Benard cells) are unable to createlocal and selective constraints to generate a proper organi-zation (understood as a disposition of parts generating aset of integrated but differentiable functions). The chemi-cal domain, however, can potentially combine dissipativeself-organization with the power of self-assembly that thestructural stability of complex molecules permits (e.g.,proteins). It is, therefore, in the (bio-)chemical domainwhere more complex forms of organization appear, inwhich a dissipative global organization generates and sus-tains functionally differentiated parts (membranes, cata-lysts, membrane gates, flagella, etc.).




5 Although conceptually separable interactive and construc-tive processes might be (and originally are) instantiated bythe same organization.

6 When using the term “environment” we are not referringonly to a physical environment but also to an ecologicaland social environment that is the result of evolutionary–ecological relationships, intraspecies social relationships,and the very recursive effect of the agent on these socio-ecological environments.

7 Although necessary, this type of organization is not suffi-cient for a minimal characterization of living beings. Liferequires that a set of hereditary and non-reactive compo-nents be coupled with the dynamic internal medium of thebasic autonomous organization (i.e., metabolism). Thesehereditary components support an open-ended capacity toconserve, instruct, modify, and reproduce organizationalcomplexity; in other words, inserting basic autonomousorganizations in a historical and collective process of evo-lution (Ruiz-Mirazo, Peretó, & Moreno, 2004).

8 We distinguish between normativity, regulation, and nor-mative regulation. From an organizational and naturalistperspective, normativity is a somehow transcendentalproperty defined by the organizational conditions of possi-bility for a dissipative organization. Regulation, however,refers to the control or active compensation of perturba-tions according to a given goal state or rule. An organismmight or might not be able to regulate itself according toits autonomously defined norms. For instance, it could beregulating its temperature to 42 ºC, maintaining this tem-perature invariant in the face of perturbations, althoughthis temperature might be harmful. Thus, it could happen(and it often happens) that an organism is regulating itselfbadly, but still regulating. So there must be a principle bywhich, independently of the actual regulatory functioningof the organism, we can justify the claim that the organismis doing it wrong. This right/wrong, good/bad, value attri-bution must be naturalized in the normativity that the FFEorganization that an autonomous system brings forth. Wecall normative regulation the regulation that is carried outaccording to the normativity of the organism.

9 The term “teleology” here should be devoid of any inten-tional or representational character, it simply denotes the par-ticipation or contribution of a process into a self-maintainingorganization. The term “teleology” was introduced by Kant,in his Critique of Judgement, to refer to such a process typeas parts are means and ends of themselves through their par-ticipation on a holistic organization; see Weber and Varela(2002) for a further development of this issue.

10 When referring to basic autonomy in living beings (uni-cellular, multicellular, with or without specialized organs,etc.) we mean the network of processes that constitute theself-maintaining and self-constructing organization of thesystem (excluding other biological functions, such as repro-

duction, immune defense, etc.). Minimal autonomy, incontrast, refers to the minimal organization capable ofrecursive self-maintenance and construction. In the mostsimple autonomous systems, interactive processes do notachieve the status of cycles. Functional actions exertedon the environment, such as ion-pumping, do not feedback through the environment, to generate additionaltransformation that in turn give rise to subsequentactions.

11 The term “constraint” is here used to mean that some physi-cal or chemical processes are limited or shaped, channeledtoward directions that do not follow from the effect of phys-ical laws or principles without the presence of that limita-tion that is in turn produced by the systemic organization(Pattee, 1972). It is here that the conceptual model byKauffman (2000) of autonomous systems as those capableof instantiating work–constraint cycles provides an insight-ful approach. To perform work (i.e., useful directedrelease of energy) constraints need to be built by the sys-tem and, conversely, to create constraints work is required.It is the characteristic coupling between exergonic (free-energy releasing and thermodynamically spontaneous) andendergonic (free-energy requiring and non-spontaneous)chemical reactions that provides the means to achieve awork–constraint closure in basic autonomous systems. Itis in this precise sense in which the system can be said tobe the active source of a functional interaction, when thesework–constraint cycles are recruited or mobilized to regu-late or direct system–environment interactions.

12 Structural stability can be thought of as a mechanism toachieve robustness that typically involves the recovery ofa stable state that was characteristic of the system beforethe perturbation took place. Jen (2002) proposes the fol-lowing working definition: “Loosely speaking, a solution(meaning an equilibrium state) of a dynamical system issaid to be stable if small perturbations to the solutionresult in a new solution that stays ‘close’ to the originalsolution for all time. Perturbations can be viewed as smalldifferences effected in the actual state of the system: thecrux of stability is that these differences remain small forall time. … A dynamical system is said to be structurallystable if small perturbations to the system itself result in anew dynamical system with qualitatively the same dynam-ics. Structural stability will typically involve a systemconfiguration such that small changes in parameters leavethe system behavior almost unaffected and/or such thatchanges in variables are compensated through negativefeedback loops that bring the system to the original stableattractor (i.e., the spontaneous compensation of perturba-tions when these fall into the basins of attraction).” (p. 2)

13 In terms of number and variety of components and inter-relations, as well as in the degree of integration of the dif-ferent parts of the system.




14 Structurally stable autonomous systems could have appearedon Earth before the invention of organisms endowed with aninstructed metabolism (a genetic code). However, it is diffi-cult to imagine how an adaptive control mechanism couldspontaneously arise and adequately function in relationto metabolic needs without natural selection assuringthat the coupling between the control mechanism and thecontrolled processes be functional for the system as awhole.

15 This is of course a simplified version of the real mecha-nism. More than activated, the expression of Lac operongene complex (which encodes three enzymes required formetabolizing lactose) is repressed. When lactose is presentin the environment, it binds to the repressor (inducing aconformational change that unlocks is repressive capacity)and transcription of Lac operon starts (but at a very lowrate). It is under the absence of glucose when an activatorprotein complex (CAP–CAMP) increases the transcriptionof an activator protein that the transcription of the Lacoperon increases.

16 It has been argued that plants possess epithelial cells,which can be sensitive to local chemical or tactile stimulitriggering a change of electric potential capable, in princi-ple, of producing fast agential responses (Simons, 1981).However, plant intercellular communication is not based onepithelial cell communication, which lacks directional andselective propagation and is unable to organize modulationand regeneration of signals. Instead, the communicationsystem of plants is based on a type of channel called plas-modesmata, which works by transporting (either passive oractively) a large variety of chemical signals. However, thismechanism is far from showing the speed, plasticity, andrecursive modulation of signals of neural networks. Not sur-prisingly, plasmodesmatal connections seem to be limited toadjacent cells (Trewavas, 2003). In addition, the bodyplan ofplants does not allow them to develop musculoskeletalstructures, which by virtue of their capacity to channelenergy into reversible mechanical motion, are of fundamen-tal importance for behavioral agency.

17 The neuron is a cell specialized in connecting sensorimo-tor surfaces in a plastic, fast, and (metabolically speak-ing) cheap way, coordination systems based on epithelialconduction being a limited precursor of neural coordina-tion.

18 None the less, it must be noted that to the hierarchicaldecoupling of the NS follows a global coupling to somemetabolic states of the organism in order to satisfy itsadaptive/metabolic needs.

19 Not in the sense of representational or semantic but in thesense of propagation of dynamic variability as measuredby information theory.

20 What one of the authors has formerly called “neuralagency” (Moreno & Etxeberria, 2005).

21 Certain forms of unicellular motility (such as the case ofchemotaxis in E. coli) can also be included under this cat-egory, despite their limited capacity to achieve relativelycomplex forms because of the intrinsic bottleneck we havepreviously described.

22 By innate, we do not mean here that there is a geneticallydetermined architecture of neural pathways in the NS butthat, given an evolutionarily stable environment, a devel-opmental process triggered by environmental and geneticfactors, certain anatomical structures are stabilized as aresult of recurrent interactions in the developmental proc-ess.

23 For instance, the nematode Caenorhabditis elegans con-tains precisely 302 neurons and about 5,000 synapses withhighly stereotyped connections whose complete wiring dia-gram is already well known and equivalent among individu-als of the same species.

24 Situated and autonomous robotics (Brooks, 1991; Clark,1997; Maes, 1990; Mataric, 2002; Pfeifer & Scheier,1999) has provided a set of insightful models of embod-ied and interactively self-organized behavior: obstacleavoidance (Brooks, 1990), wall following (Steels, 1991),behavioral categorization (Cliff, Husbands, & Harvey,1993) and a number of other interactive behavioral phe-nomena that exploit recurrent interactions with the envi-ronment (Ziemke, 2003).

25 By this term we mean not only metabolic self-mainte-nance but a number of organizational constraints derivedfrom the evolutionary dimension of living beings (repro-duction, kin caring, sexual selection, etc.).

References

Ashby, W. R. (1952). Design for a brain. The origin of adap-tive behavior. London: Chapman & Hall.

Barandiaran, X. (2004). Behavioral adaptive autonomy. Amilestone in the Alife route to AI? In J. Pollack, M.Bedau, P. Husbands, T. Ikegami, & R. Watson (Eds.),Proceedings of the 9th International Conference on Artifi-cial Life (pp. 514–521). Cambridge, MA: MIT Press.

Barandiaran, X. (2007). Mental life: Conceptual models andsynthetic methodologies for a post-cognitivist psychology.In B. Wallace, A. Ross, J. Davies, & T. Anderson (Eds.),The world, the mind and the body: Psychology after cogni-tivism (pp. 49–90). London: Imprint Academic.

Barandiaran, X., & Moreno, A. (2006). On what makes certaindynamical systems cognitive. Journal of Adaptive Behav-ior, 14(2), 171–185.

Bickhard, M. H. (1993). Representational content in humansand machines. Journal of Experimental and TheoreticalArtificial Intelligence, 5, 285–333.




Brooks, R. A. (1990). Elephants don’t play chess. In P. Maes(Ed.), Designing autonomous agents (pp. 3–15). Cam-bridge, MA: MIT Press.

Brooks, R. A. (1991). Intelligence without representation. Arti-ficial Intelligence Journal, 47, 139–160.

Christensen, W. D., & Bickhard, M. H. (2002). The processdynamics of normative function. Monist, 85(1), 3–28.

Clark, A. (1997). Being there. Putting brain, body, and worldtogether again. Cambridge, MA: MIT Press.

Cliff, D., Husbands, P., & Harvey, I. (1993). Evolving visuallyguided robots. In J.-A. Meyer, H. Roitblat, & S. Wilson(Eds.), From Animals to Animats 2: Proceedings of the2nd International Conference on Simulation of AdaptiveBehavior (SAB92) (pp. 374–383). Cambridge, MA: MITPress/Bradford Books.

Collier, J. (1999). Autonomy and process closure as the basis forfunctionality. In J. L. R. Chandler & G. van der Vijver(Eds.), Closure: Emergent organizations and their dynam-ics. Annals of the New York Academy of Sciences, 901(1),280–290.

Di Paolo, E. (2003). Organismically inspired robotics. In K.Murase & T. Asakura (Eds.), Dynamical systems approachto embodiment and sociality (pp. 19–42). Adelaide, Aus-tralia: Advanced Knowledge International.

Di Paolo, E. (2005). Autopoiesis, adaptivity, teleology, agency.Phenomenology and the Cognitive Sciences, 4(4), 429–452.

Edelman, G. M. (1987). Neural Darwinism: The theory ofgroup neuronal selection. New York: Basic Books.

Izquierdo-Torres, E., & Di Paolo, E. A. (2005). Is an embodiedsystem ever purely reactive? In M. S. Capcarrere, A. A.Freitas, B. J. Bentley, C. G. Johnson, & J. Timmis (Eds.),Advances in Artificial Life. Proceedings of the Eighth Euro-pean Conference (ECAL 2005), Lecture Notes in ComputerScience series (pp. 252–261). Berlin: Springer Verlag.

Jen, E. (2002). Stable or robust? What is the difference?Retrieved 22 April 2008, from http://www.santafe.edu/~erica/stable.pdf

Kauffman, S. (2000). Investigations. Oxford: Oxford Univer-sity Press.

Llinás, R. R. (2001). The I of the vortex. From neuron to self.Cambridge, MA: MIT Press.

Maes, P. (Ed.). (1990). Designing autonomous agents. Theoryand practice from biology to engineering and back. Cam-bridge, MA: MIT Press.

Mataric, M. J. (2002). Situated robotics. In Encyclopedia ofCognitive Science (Vol. 4, pp. 25–30). London: NaturePublishing Group, Macmillan Reference Limited.

Maturana, H., & Varela, F. (1980). Autopoiesis and cognition:The realization of the living. Dordrecht: Reidel.

McGeer, T. (1990). Passive dynamic walking. InternationalJournal of Robotics Research, 9(2), 62–82.

Millikan, R. G. (1984). Language, thought and other biologicalcategories. Cambridge, MA: MIT Press.

Moreno, A., & Barandiaran, X. (2005). A naturalized accountof the inside–outside dichotomy. Philosophica, 73, 11–26.

Moreno, A., & Etxeberria, A. (2005). Agency in natural andartificial systems. Artificial Life, 11(1–2), 161–176.

Moreno, A., Etxeberria, A., & Umerez, J. (2008). The auton-omy of biological individuals and artificial models. Bio-Systems, 91(2), 309–319.

Moreno, A., & Lasa, A. (2003). From basic adaptivity to earlymind. Evolution and Cognition, 9(1), 12–30.

Moya, A., Peretó, J., Gil, R., & Latorre, A. (2008). Learninghow to live together: Genomic insights into prokaryote–animal symbiosis. Nature Reviews Genetics, 9, 218–229.

Nicolis, G., & Prigogine, Y. (1977). Self-organization in non-equilibrium systems. New York: Wiley.

Pattee, H. H. (1972). Laws and constraints, symbols and lan-guages. In C. H. Waddington (Ed.), Towards a theoreticalbiology 4, Essays (pp. 248–258). Edinburgh: EdinburghUniversity Press.

Pfeifer, R., & Scheier, C. (1999). Understanding intelligence.Cambridge, MA: MIT Press.

Ruiz-Mirazo, K. (2001). Physical conditions for the appear-ance of autonomous systems with open-ended evolution-ary capacities. Ph.D. dissertation, University of theBasque Country (complete version only available in Span-ish at http://www.sc.ehu.es/PhBioGroup).

Ruiz-Mirazo, K., & Moreno, A. (2004). Basic autonomy as afundamental step in the synthesis of life. Artificial Life,10, 235–259.

Ruiz-Mirazo, K., Peretó, J., & Moreno, A. (2004). A universaldefinition of life: Autonomy and open-ended evolution.Origins of Life and Evolution of the Biosphere, 34, 323–346.

Seth, A. K., & Edelman, G. M. (2004). Environment andbehavior influence the complexity of evolved neural net-works. Adaptive Behavior, 12(1), 5–21.

Shigenobu, S., Watanabe, H., Hattori, M., Sakaki, Y., &Ishikawa, H. (2000). Genome sequence of the endocellu-lar bacterial symbiont of pahids Buchenera sp. APS.Nature, 407, 81–86.

Simons, P. J. (1981). The role of electricity in plant move-ments. New Phytologist, 87, 11–37.

Steels, L. (1991). Towards a theory of emergent functionality.In J. Meyer & R. Wilson (Eds.), Simulation of adaptivebehavior (pp. 451–461). Cambridge, MA: MIT Press.

Tononi, G., Edelman, G. M., & Sporns, O. (1998). Complexityand coherency: Integrating information in the brain.Trends in Cognitive Science, 2(12), 474–484.

Trewavas, A. (2003). Aspects of plant intelligence. Annals ofBotany, 92, 1–20.

van Duijn, M., Keijzer, F. A., & Franken, D. (2006). Principlesof minimal cognition. Journal of Adaptive Behavior, 14(2),157–170.




Weber, A., & Varela, F. J. (2002). Life after Kant: Natural pur-poses and the autopoietic foundations of biological indi-viduality. Phenomenology and the Cognitive Sciences, 1,97–125.

Ziemke, T. (2003). Embodied AI as science: Models of embod-ied cognition, embodied models of cognition, or both? In F.Iida, R. Pfeifer, L. Steels, & Y. Kuniyoshi (Eds.), Embodiedartificial intelligence (pp. 27–36). Heidelberg: Springer-Verlag.

About the Authors

Xabier Barandiaran is a Ph.D. student at the University of the Basque Country. He is agraduate student in philosophy from the University of Deusto (Bilbao, Spain) andobtained an M.Sc. in evolutionary and adaptive systems at the University of Sussex(Brighton, UK). His main research areas are the philosophy of artificial life, neurophiloso-phy, enactive cognitive science, and naturalized epistemology.

Alvaro Moreno is Professor of Philosophy of Science at the University of the BasqueCountry, in Spain, where he has created a research group specialized in complex sys-tems, philosophy of biology, and artificial life. He is the author of over a hundred publica-tions and organizer of several international workshops on the relationship betweenbiology and cognition. Address: Departamento de Logica y Filosofia de la Ciencia, Uni-versidad del Pais Vasco UPV-EHU, Apartado 1249, 20080 San Sebastian, Spain. E-mail: [email protected]



barandarian, x. moreno, a. (2008) adaptivity. from metabolism to behavior.pdf

Documents