Published in Margaret Boden, ed., 1996, The Philosophy of Artificial Life (Oxford University Press), pp. 332-357. Missing two figures.

The philosopher tries to define it [life], but no definition will cover its infinite and self-contradictory variety.

J. B. S. Haldane (1937: 64)

 

The idea of life, the sense of being alive, are the most familiar and the most difficult to understand of the concepts we meet.

J. Lovelock (1988: 16)

 

Few biologists today think it is worthwhile to pay much attention to that distinction [between life and non-life].

C. Taylor (1992: 26)

 

The nature of life used to concern philosophers&endash;think of Aristotle and Kant&endash;but most philosophers ignore the issue today, perhaps because it seems too "scientific". At the same time, most biologists also ignore the issue, perhaps because it seems too "philosophical". But this intellectual context has just been changed by the advent of the new interdisciplinary field of artificial life. Using the resources of both philosophy and artificial life, I will try to revive the question of the nature of life and defend a certain answer to it.

1. Facts and puzzles about the phenomena of life.

Life is amazing. It is all around us in a diversity of forms, ranging from microscopic bacteria to ancient towering trees, from almost inert lichen to transient insect blooms, from birds flocking in the sky to thriving colonies of tube worms at inky deep-sea vents. The first forms of life on earth spontaneously arose out of a preexisting prebiotic chemical soup. From those simple origins has evolved a diverse hierarchy of forms of life, which includes the most complex objects in the known universe. Individual living entities (organisms) maintain their self-identity and their self-organization while continually exchanging materials and energy and information with their local environment. Different species of life flexibly and tenaciously exploit various niches in the environment. When viewed on a long enough time scale, life forms are always changing, adjusting, producing novel responses to unpredictable contingencies, adapting and evolving through blindly opportunistic natural selection.

Not all the diversity and complexity and change in life is adaptive, of course. Random drift, architectural constraints and other non-adaptive factors have their influence. But what is especially distinctive and striking about life in the long run is the supple, open-ended evolutionary process that perpetually produces novel adaptations. In fact, I will contend in this paper that supple adaptation defines life at its most general.

There are plenty of puzzles about the concept of life. The concrete objects ready to hand are usually easily classified as living or non-living. Fish and ants are alive while candles, crystals and clouds are not. Yet many things are genuinely puzzling to classify as living or not. Viruses are one borderline case, biochemical soups of evolving RNA strings in molecular genetics laboratories are another. The Gaia hypothesis (Lovelock 1988), according to which the entire chemical and biological environment around the surface of the earth (including things like the oceans and the atmosphere) constitute one living organism, also strains the ordinary concept of life. So does the search for extraterrestrial life. Extraterrestrial life forms, if any exist, might well not depend on DNA-encoded information or, indeed, any familiar carbon chemistry processes. How would we recognize extraterrestrial life if we found it? We have no reason to suppose it will have any of the accidental characteristics found in familiar forms of life. What, then, are the essential properties possessed by all possible forms of life? The search for extraterrestrial life needs some answer to this question, for we can search for life only if we have a prior conception of what life is.

The phenomena of life raise a variety of subtle and controversial questions. Borderline cases like viruses raise the general issue of whether life is a black-or-white property, as it seems at first blush, or whether it comes in shades of gray. Early life forms somehow originated from pre-biotic chemical soup. Does this imply that there is an ineliminable continuum of things being more or less alive, as many suppose (e.g., Cairns-Smith 1985, Küppers 1985, Bagley and Farmer 1992, Emmeche 1994, Dennett 1995)? Another subtle question concerns the different levels of living phenomena&endash;such as cells, organs, organisms, ecosystems&endash;and asks in what senses (if any) the concept of life applies at these various levels. Mayr (1982) seems to be especially sensitive to this question, although he has no ready answer. Recently a third question has been receiving lots of attention (e.g., Langton 1989a, Emmeche 1992): Does the essence of life concern matter or form? On the one hand, certain distinctive carbon-based macromolecules play a crucial role in the vital processes of all known living entities; on the other hand, life seems to be more in the nature of a process than a kind of substance. The relationship between life and mind raises a fourth question. When we consider plants, bacteria, insects, and mammals, for example, we apparently find different kinds of mental activity, and it seems that different degrees of behavioral sophistication correspond to different levels of intelligence. Might the various forms of life and mind be somehow connected?

To answer questions like these and make sense of the puzzling phenomena of life, we need a sound and compelling grasp of the nature of life. Can any property embrace and unify not only life's existing diversity but also all its possible forms? What is the philosophically and scientifically most plausible way to account for the characteristic life-like features of this striking diversity of phenomena? How can we resolve the controversies about life? The concept of life as supple adaptation, explained below, is my attempt to address these issues.

Notice that our ordinary, everyday concept of life does not settle what the true nature of life is. Thus, we are not concerned here with careful delineation of the paradigms and stereotypes that we commonly associate with life. We want to know what life is, not what people think life is. Glass does not fall under the everyday concept of a liquid, even though chemists tell us that glass really is a liquid. Likewise, we should not object if the true nature of life happens to have some initially counterintuitive consequences.

2. Conceptions of life.

My main goal in what follows is to develop one particular conception of life&endash;life as supple adaptation. But it will be useful first to outline some other conceptions of life in order to introduce some of the competition. I will focus on three prominent kinds of views: life as a loose cluster of properties, life as a specific set of properties, and life as metabolization. There are a number of other interesting accounts of life, such as those based on self-replication (Poundstone 1985), autopoiesis (Maturana and Varela 1973), and closed causal loops (Rosen 1991), but this is not the occasion to discuss them all.

A skeptic might question whether there is any interesting single, all-inclusive account of life. The demise of vitalism taught us that no non-physical substance or force is distinctive of all instances of life. The skeptic asks what guarantees that some single property is distinctive of the unified diversity of life. For all we know, the truth about life might be no more unified than a collection of overlapping properties from overlapping disciplines, such as population genetics, molecular genetics, evolution, ecology, cytology, biochemistry, and physiology. Such skeptics often argue that life is characterized merely by a cluster of loosely connected properties. The individual properties in the cluster are held to be typically but not necessarily possessed by living entities; the diversity of living phenomena are thought to bear only a Wittgensteinian family resemblance.

A number of such clusters have been proposed. Farmer and Belin (1992: 818), for example, list eight characteristics: process; self-reproduction; information storage of self-representation; metabolization; functional interactions with the environment; interdependence of parts; stability under perturbations; and the ability to evolve. Farmer and Belin (1992: 818) explain that what drives them to a cluster conception of life is their despair of writing down anything more precise than "a list of properties that we associate with life" since

[t]here seems to be no single property that characterizes life. Any property that we assign to life is either too broad, so that it characterizes many nonliving systems as well, or too specific, so that we can find counter-examples that we intuitively feel to be alive, but that do not satisfy it.

Taylor provides a similar justification for characterizing life with a similarly loosely linked list of properties: "Each property by itself, even when considered with others, is unable to clearly delineate the living from the non-living, but together they do help to characterize what makes living things unique" (1992: 26).

There is a special virtue in viewing life as a loose cluster of properties, for this provides a natural explanation of why life has vague boundaries and borderline cases. The main drawback of cluster conceptions is that they inevitably make life seem rather arbitrary or, at least, mysterious. A cluster offers no explanation of why that particular cluster of properties is a fundamental and ubiquitous natural phenomenon. We must acknowledge that there is no a priori guarantee of a single, unifying account of life; such are the possible hazards of philosophy and science. Still, there might be such a account; such are the possible fruits of philosophy and science. To settle this question we must seek and evaluate more unified explanations of life. A cluster conception is a fall-back position that can be justified only after all candidate unified views have failed.

Life is sometimes characterized by a list of properties that are intended to characterize not a family resemblance but something much closer to necessary and sufficient conditions. Most lists are relatively short, and most contain many of the same properties. Monod (1971) lists three defining characteristics of life: "teleonomic" or purposeful behavior, autonomous morphogenesis, and reproductive invariance. Crick (1981) focuses on a somewhat related set: self-reproduction, genetics and evolution, and metabolization. Crick's list is almost identical with Küppers's (1985): metabolism, self-reproduction, and mutability. In The Problems of Biology Maynard Smith (1986) cites two properties: metabolism and parts with functions. Ray (1992) cites two others: self-reproduction and the capacity for open-ended evolution. An especially comprehensive list is produced by Mayr (1982: 53), who thinks that "[t]he process of living . . . can be defined" by a list of "the kinds of characteristics by which living organisms differ from inanimate matter". It is worthwhile summarizing Mayr's entire list:

1. All levels of living systems have an enormously complex and adaptive organization.

Mayr's list is a quite useful characterization of the special hallmarks of living systems, and it cannot help but deepen our sense of wonder and perplexity about what root cause could conspire to make this striking collection of features present in such an indefinite diversity of natural phenomena. In this, it suffers from the same weakness that besets cluster conceptions of life. We want an account of why these properties all coexist. Rather than settling this question, the list raises it. Of course, as our skeptic will remind us, the list of features might have no "root cause". The list might be refer to something like a medical syndrome&endash;a collection of symptoms with no underlying cause. But when doctors discover the characteristic coexistence of a list of symptoms, they seek an underlying cause, and they sometimes find an underlying cause for what previously seemed to be merely a syndrome. In the same way, it is appropriate always to keep one's mind open to the possibility of finding an underlying cause for any conjunction of hallmarks of life.

Schrödinger proposed persisting in the face of the second law of thermodynamics, by means of the process of metabolization, as the defining feature of life. The following passages outline his position (1969: 74-76):

What is the characteristic feature of life? When is a piece of matter said to be alive? When it goes on "doing something", moving, exchanging material with its environment, and so forth, and that for a much longer period than we would expect an inanimate piece of matter to "keep going" under similar circumstances. . . . It is by avoiding the rapid decay into the inert state of "equilibrium" that an organism appears so enigmatic; . . . How does the living organism avoid decay? The obvious answer is: By eating, drinking, breathing and (in the case of plants) assimilating. The technical term is metabolism. . . . [T]he essential thing in metabolism is that the organism succeeds in freeing itself from all the entropy it cannot help producing while alive.

It is compelling to think that life centrally involves the process of metabolization. For one thing, this nicely explains our intuition that a crystal is not alive (there is a metabolic flux of molecules only at the crystal's edge, not inside it). But perhaps what is most compelling is that any possible form of life that persists in the face of the second law of thermodynamics apparently must have a metabolization (or, as Schrödinger puts it, must feed upon negative entropy). By this argument, metabolization is at least a necessary condition of all physical forms of life.

The main drawback of metabolization as an all-encompassing conception of life is that intuitively many metabolizing entities seem not to be alive and not to involve life in any way. Standard examples include a candle flame, a vortex, and a convection cell (Maynard Smith 1986, Bagley and Farmer 1992). These examples by themselves do not prove conclusively that metabolization is not sufficient for life. An adequate and attractive conception of life need not classify as alive all and only those things we intuitively (even confidently) classify as alive. What matters is whether metabolization is the distinctive feature that in fact explains the unified diversity of life. Any convincing defense of life as metabolization must show that metabolization accounts for life's most characteristic features; this will thereby establish that candles and the like share with all life forms what is essential to life. Whether this is possible seems doubtful.

3. Life as supple adaptation.

It is sometimes suggested that the central feature underlying all life is the evolutionary process of adaptation. What is emphasized is sometimes the blind operation of natural selection, sometimes the general process of evolution, and sometimes the adaptive traits produced by these means. In each case the central idea is that what distinguishes life is an underlying automatic and open-ended capacity to adapt appropriately to unpredictable changes in the environment. From this perspective, what is distinctive of life is the way in which adaptive evolution automatically fashions new and intelligent strategies for surviving and flourishing as local contexts change.

In The Theory of Evolution Maynard Smith (1975: 96f) succinctly explains the justification for this view that life crucially depends on the evolutionary process of adaptation:

We shall regard as alive any population of entities which has the properties of multiplication, heredity and variation. The justification for this definition is as follows: any population with these properties will evolve by natural selection so as to become better adapted to its environment. Given time, any degree of adaptive complexity can be generated by natural selection.

Cairns-Smith (1985: 3) also emphasizes adaptive evolution's central role in accounting for life's characteristic features:

[N]atural selection is only one component of the mechanism of evolution. Any theory that is to explain the variety and complexity of living things must also take into account the varied and varying challenges set up by a varied and varying environment. Nature, as breeder and show judge, is continually changing her mind about which types should be awarded first prize: changing selection pressures have been a key part of her inventiveness.

But nevertheless natural selection has been the key component, the sine qua non. Without it, living things could not even stay adapted to a given set of circumstances, never mind become adapted to new ones. Without natural selection the whole adventure would never have got off the ground. That kind of in-built ingenuity that we call 'life' is easily placed in the context of evolution: life is a product of evolution. [emphasis in original]

These remarks suggest why it is plausible that the process of adaptive evolution could explain all of life's hallmarks. Although I will not review this exercise here, it is easy enough to see how this approach to life can mount plausible explanations of all of the entries on Mayr's list of the distinctive features of life.

I endorse a specific version of this approach to life. I will content myself with developing the general form of my view and defending it against various objections. Since the notion of adaptive evolution plays a pivotal role in my account, I will focus most of my attention on clarifying the relevant sense of adaptation and illustrating how it can be given a precise, quantitative, empirical explication.

Evolving systems that (according to this approach) underlie living phenomena involve a specific form of adaptation. The systems are automatically evolving in an open-ended manner and thereby continually producing new adaptive traits. The essential principle that explains the unified diversity of life seems to be this suppleness of the adaptive processes&endash;its unending capacity to produce novel solutions to unanticipated changes in the problems of surviving, reproducing, or, more generally, flourishing. Some forms of adaptation are rigid, such as those exemplified by artifacts like street lights or thermostats which have strictly limited options: turn on or turn off. By contrast, supple adaptation involves responding appropriately in an indefinite variety of ways to an unpredictable variety of contingencies. (The contingencies are "unpredictable" in the sense that the living system has no chance of already adapting to them.) Phrases like "open-ended evolution" (Lindgren 1992: 310; Ray 1992: 372) or "perpetual novelty" (Holland 1992: 184) have also been used to refer to this same process.

One might think that natural selection will inevitably produce supple adaptation, but this is wrong. When selection is made on the basis of a fixed fitness function, the resulting adaptive dynamics eventually stabilize rather than continually produce adaptive novelty. For example, Mitchell and Forrest (1994) explain that adaptation toward a fixed goal is the characteristic&endash;and desired&endash;outcome when natural selection is implemented in a so-called "genetic algorithm" and applied to engineering problems such as optimization (e.g., circuit design and job shop scheduling), automatic programming (evolving computer programs for specific tasks like sorting lists), and machine learning (e.g., predicting protein structure). Thus, natural selection will yield supple adaptation only if the criteria for selection change as the system evolves. Changes in selection criteria can be driven by an independently changing environment, but this is not necessary. A significant aspect of the environment to which any given organism must adapt is all the other organisms with which it interacts. So, when a given organism adapts and changes, the evolutionary context of all the other organisms changes. Thus, even without an externally changing environment, adaptation can be a co-evolutionary process that internally changes the selection pressures which shape adaptation, thus making open-ended adaptive evolution an intrinsic property of the system (Packard 1989, Holland 1992). And this intrinsic form of supple adaptation probably happens even when external factors also change the environment.

There are different ways to define life in terms of supple adaptation, and I want to emphasize two special aspects of my own approach. First, I want to say that supple adaptation does not merely produce living entities. In addition, the entity that is living in the primary sense of that term is the supplely adapting system itself. Other entities that are living are living in a secondary sense by virtue of bearing an appropriate relationship to a supplely adapting system. Different kinds of living entities (organisms, cells, etc.) will stand in different kinds of relationships to the supplely adapting system from which they derive their life, although these relationships in general all involve ways in which the adapting system generates and sustains (in the "right way") the entity. So, my conception of life can be captured by something with the form roughly of the following three definitions.

A. X is living iff X is living1 or X is living2.

B. X is living1 iff X is a system undergoing supple adaptation.

C. X is living2 iff there is some living1 system Y such that either (1) X meets condition A1 & Y meets condition B1 & X bears relation C1 to Y, or (2) X meets condition A2 & Y meets condition B2 & X bears relation C2 to Y, or . . . or (n) X meets condition An & Y meets condition Bn & X bears relation Cn to Y.

Definitions A &emdash; C indicate only the general form of my approach, which I will not try here to delineate more precisely. I also will not try here to specify the content of the clauses in definition C. My intention is to show only how my approach construes supplely adapting systems as the primary living entities.

The second respect in which my view may differ from other conceptions of life as supple adaptability concerns the difference between a capacity and its exercise. Whereas some might refer only to a system's capacity to undergo supple adaptation, I hold that life involves the exercise of this capacity. For me the key is not supple adaptability but supple adaptation. I don't want to over emphasize this difference, for the same process is implicated in both cases. Furthermore, a system that is actually undergoing supple adaptation is not continuously adapting; the adaptation often happens in fits and starts. A system is exhibiting supple adaptation over a given time period provided the quiescent periods without adaptation are not permanent, i.e., every quiescent period is followed by the evolution of new adaptations. Nevertheless, if a system could undergo supple adaptation never does, then I would be inclined to say that it could be living and could support life but isn't.

With this background, we can start to see how my view of life as supple adaptability could respond to a number of criticisms. For example, mules, the last living member of an about-to-be extinct species, neutered and spayed animals are all alive, but being infertile such entities cannot play any role in the supple adaptation of their own lineage or larger population. Thus, one might think that infertile organisms are counterexamples to the necessity of the my account of life. However, these infertile organisms exist only because of their connections with other, fertile organisms which do play an active role in a biosphere that undergoes supple adaptation.

One might worry that an evolving system's supple adaptation has the wrong "logical form" to be a concept of life. This worry starts with the idea that individual organisms are the entities that are alive, then observes that the whole evolving population of organisms is necessarily of a different logical category than an individual organism, and so concludes that life cannot consist in a population undergoing supple adaptation. However, this objection has no force for those who are seeking the fundamental explanation of the diversity of living phenomena. Supple adaptation could provide this explanation even though an individual living organism is itself only a small and transitory part of the whole adapting population.

One might worry that this approach to life would have difficulty establishing without circularity that an organism's corpse is not alive, for the corpse and the living organisms are both produced by the adapting population in the same way. However, the corpse and living organism differ in various intrinsic properties like metabolization. Although I have not tried to fill in the content of the clauses in definition C above, it would contain some clause&endash;call it clause (i)&endash;applying to living organisms and conditions like undergoing metabolization would presumably play the role of Ai in clause (i). In this way the definition would non-circularly demark living organisms.

The possibility of an ecology that has reached a state of stable equilibrium and altogether stopped adapting might seem to provide a more direct challenge to my view of life. After all, the organisms in such so-called "climax" ecosystems are certainly alive, yet the ecosystem containing them is not undergoing supple adaptation, so these organisms would seem to fall outside my definition. However, this problem vanishes when one adopts a sufficiently broad perspective. Not only do climax ecosystems originate through a process of supple adaptation, they all eventually break out of their state of no adaptation. The quiescent periods of stability are transitory; in the long run supple adaptation always occurs.

Things that exhibit open-ended adaptation but seem devoid of all life might seem to be counterexamples to the sufficiency my view of life. Viruses are unquestionably adapting to all our best efforts to eradicate them&endash;the AIDS virus apparently does this with remarkable rapidity&endash;yet viruses are a classic case on the borderline of life. Even populations of the tiny clay crystallites that make up mud seem to have the flexibility to adapt and evolve by natural selection (Cairns-Smith 1985, Bedau 1991), and so do autocatalytic networks of chemical species (Bagley and Farmer 1992), yet to our ordinary way of thinking evolving populations of crystals or chemicals involve no life at any level. In addition, human intellectual activities and economic markets both can seem to have a flexible, open-ended capacity to adapt to unpredictably changing circumstances; empirical support for such conjectures might be found by analyzing, for example, citation patterns in the Science Citation Index, traffic patterns on the internet, or patterns of transactions in stock exchanges. While intellectual and economic activities are generated by living creatures, the evolving intellectual or economic systems themselves might seem quite unlike anything that we would ordinarily want to call living. In fact, cases like these have prompted the proposal that supple adaptation is only a necessary condition of systems containing living entities (Bedau and Packard 1992). However, since here I am not offering supple adaptation as an explication of our everyday concept of life, unintuitive classification of some cases is no fatal flaw. These counterintuitive cases do not refute the hypothesis that supple adaptation is the underlying explanatory factor that unifies the diverse phenomena of life. If this hypothesis is true, and if populations of viruses and clay crystallites, autocatalytic networks of chemicals, and even human intellectual and economic systems exhibit supple adaptation, they all well deserve to be thought of as "living" for they all depend on the same underlying process. Our ordinary language already sanctions this conclusion to a degree, since we find it quite natural and compelling to speak of the "vitality" of virus populations and intellectual and economic systems.

It is easy to conceive of circumstances that violate my account of life. Nothing prevents us from entertaining the scientific fantasy of species that never evolve and adapt, as creationists perhaps suppose. For all I know, this is logically possible. So is the possibility that all organisms were created in seven days by an omnipotent, omniscient, and omnibenevolent deity. So is the possibility that there has been and ever will be only exactly one living organism. But I take it that these fantasies are just that&endash;fantasies, with no bearing on the true nature of any form of life that we could discover or synthesize. My claim&endash;a posteriori, perhaps, but still true, I wager&endash;is that all living organisms ultimately derive their existence and their characteristic life-like features from having the right sort of connection with a system undergoing supple adaptation.

Our final evaluation of the conception of life as supple adaptation depends on learning more about supple adaptation. We need answers to at least the following questions:

1. How can supple adaptation be empirically observed and measured?

2. What minimal models exhibit supple adaptation?

3. What fundamental laws characterize supple adaptation?

In my estimation, these are some of the most fundamental unresolved issues which our final judgment about the scientific and philosophical utility of the conception of supple adaptation must await.

We are not completely at a loss about how to answer these questions, though. For example, we noted above that metabolism is required for a physical entity to persist in the face of the second law of thermodynamics, so any physical system exhibiting supple adaptation must rely on metabolic processes to sustain itself. Similarly, some form of self-replication seems to be implicated in any evolutionary adaptation (Poundstone 1985, and Bagley, Farmer and Fontana 1992). Furthermore, we noted above that supple adaptation is often, and perhaps always, a co-evolutionary process that itself changes the internal selection pressures that drive adaptation. These conclusions all constrain what a minimal model of supple adaptation must be like, and thus inform the answer to question 2.

The fundamental laws (if any) that characterize supple adaptation are much less certain. A speculative but intriguing suggestion of one such law is the hypothesis that supple adaptation depends on evolutionary dynamics being neither too simple nor too chaotic but just complex enough, poised "at the edge of chaos" (see, e.g., Langton 1992, Kaneko 1993/1994, Bedau and Bahm 1994, Bedau and Seymour 1994). Whether or not this suggestion will prove to have lasting merit, it does illustrate what an answer to question 3 might look like and how it might be discovered.

As question 1 implies, a crucial requirement for improving our understanding of supple adaptation&endash;as well as our ability to assess its relevance for life in general&endash;is to devise an empirical method for measuring supple adaptation in actual systems. In the balance of this paper I sketch such a method and apply it to data from artificial life computer models. Contemporary philosophers might be unsure how to grasp and evaluate my explication of supple adaptation, because the field of artificial life is relatively new and unfamiliar (and controversial). Thus, some background about artificial life is in order. My review will be brief. A comprehensive presentation of work in the field can be found in the proceedings of recent artificial life conferences (e.g., Farmer et al. 1986, Langton 1989b, Langton et al. 1992, Varela and Bourgine 1992, Brooks and Maes 1994, Stonier and Yu 1994). Since artificial life promises to prompt and inform philosophical discussion on a variety of topics in a variety of ways (Bedau 1992, Dennett 1994), it is well worthwhile for philosophers to become better informed about the field.

4. Artificial life's emergent thought experiments.

Artificial life is an interdisciplinary field that attempts to understand the essential nature of living systems by means of devising and studying computationally implemented models of the characteristic processes of living systems. These processes include self-organization, metabolization, self-reproduction, and adaptive evolution.

Those who work in artificial life are betting on the working hypothesis that the essential nature of the fundamental processes of life can be implemented in relatively simple computer models. This working hypothesis is at odds with the conclusions often drawn from the rampant historicity, contingency, and variety of biological systems (e.g., Mayr 1988, Gould 1989). If true, the working hypothesis would be striking, for the complexity of living phenomena could then have a fundamentally simple explanation. Artificial life involves the search for such models. In the attempt to capture the simple essence of vital processes, the models abstract away from the vast majority of the details present in natural living systems. With no pretence of accurately modeling the particular features of particular natural systems, these are "idea" models designed to explore the consequences of certain simple premises.

Most natural systems exhibiting complex autonomous behavior seem to be parallel, distributed networks of communicating "agents". These agents make decisions about how to behave based on selective information about their own local environment, and their behavior directly affects only their own local environment. Following this lead, artificial life is exploring the dynamics of highly parallel models of simple agents in simple local environments.

These models are "emergent" in that they generate complex macro-level dynamics from simple micro-level mechanisms. (I use "micro" and "macro" in a generalized sense, as relative terms. For me, an entity exists at a micro-level relative to a macro-level population of similar micro-level entities. Micro-level entities need not be literally microscopic.) This form of emergence arises in contexts in which there is a system, call it S, composed out of "micro-level" parts; the number and identity of these parts might change over time. S has various "macro-level" states (macrostates) and various "micro-level" states (microstates); S's microstates are the states of its parts. S's macrostates are structural properties constituted wholly out of microstates; macrostates typically are various kinds of statistical averages over microstates. Further, there is a relatively simple and implementable microdynamic, call it D, which governs the time evolution of S's microstates; in general, the microstate of a given part of the system at a given time is a result of the microstates of "nearby" parts of the system at preceding times. Given these assumptions, I will say that a macrostate P of system S with microdynamic D is emergent if and only if P (of system S) can be explained from D, given complete knowledge of external conditions, but P can be predicted (with complete certainty) from D only by simulating D, even given complete knowledge of external conditions.

Although this is not the occasion to develop and defend this concept of emergence (see Bedau forthcoming), I should clarify three things. First, "external conditions" are conditions affecting the system's microstates that are extraneous to the system itself and its microdynamic. One kind of external condition is the system's initial condition. If the system is open, then another kind of external condition is the contingencies of the flux of parts and states into S. If the microdynamic is nondeterministic, then each nondeterministic effect is another external condition.

Second, given the system's initial condition and other external conditions, the microdynamic completely determines each successive microstate of the system. And the macrostate P is a structural property constituted out of the system's microstates. Thus, the external conditions and the microdynamic completely determine whether or not P obtains. In this specific sense, the microdynamic plus the external conditions "explain" P. One must not expect too much from these explanations. For one thing, the explanation crucially depends on the massive contingencies in the initial conditions; it is awash with accidental information about S's parts. Furthermore, the explanation might be too detailed for anyone to "survey" or "grasp". It might even obscure a simpler, macro-level explanation that unifies systems with different external conditions and different microdynamics. Nevertheless, since the microdynamic and external conditions determine P, they explain P.

Third, in principle we can always predict S's behavior with complete certainty, for given the microdynamic and external conditions we can always simulate S as accurately as we want. Thus, the issue is not whether S's behavior is predictable&endash;it is, trivially&endash;but whether we can predict S's behaviors only by simulating S. When trying to predict a system's emergent behavior, in general one has no choice but simulation. This notion of predictability only through simulation is not anthropocentric; nor is it a product of some specifically human cognitive limitation. Even a Laplacian supercalculator would need to observe simulations to discover a system's emergent macrostates.

Because based on "idea" models, artificial life simulations are in effect thought experiments&endash;but emergent thought experiments. As with the familiar "armchair" thought experiments, artificial life simulations attempt to answer "What if X?" questions. What is distinctive about emergent thought experiments is that they uncover consequences that can be discerned only by simulation. Artificial life's computational working hypothesis gives the field a characteristic and distinctive empirical character. A major part of the evidence discussed in the field is empirical evidence about the emergent properties of computer simulations, like those illustrated in the next section.

5. Measuring supple adaptation with usage statistics.

Before we can subject my account of life to sustained critical scrutiny, we must have some method for determining whether, and to what extent, an arbitrary system exhibits supple adaptation; i.e., we need an explication of supple adaptation. Below I will present such an explication in the context of artificial life models. These computer simulations enable me to illustrate my method easily and concretely, but the method is wholly general and applies in exactly the same form to data from natural systems.

Measuring supple adaptation in a computer model or even a natural system does not in itself make my account of life more plausible, of course. But my explication does show that the notion of supple adaptation can be made clear, coherent, and empirically verifiable&endash;which is a necessary condition on any compelling view of life. In addition, because the explication is quantitative, it provides a measure&endash;if my account of life is correct&endash;that can be used in comparing the degree of "life" in living systems. All of this provides a new empirical framework for investigating the nature and implications of the conception of life as supple adaptation.

In the artificial life literature one encounters reports of people having seen supple adaptation emerging in artificial life models. For example, Tom Ray says that his tierra model "has generated rapidly diversifying communities of self-replicating organisms exhibiting open-ended evolution by natural selection" (Ray 1992: 373). Lindgren (1992) has made similar claims about the iterated prisoner's dilemma with noise, as has Holland (1992) about the echo class of models. There is no question that these claims help shed some light on supple adaptation, but so far they are all just informal anecdotes using unstated and undefended criteria for what supple adaptation is. Our understanding of supple adaptation would significant increase if we had a general, objective method for empirically verifying claims of supple adaptation.

To address this issue, it is salutary to return to fundamentals and to restrict our attention (at least initially) to evolutionary contexts. At bottom, the process of adaptation is the process by which traits are selected for some beneficial effect or function that they perform; such traits are called "adaptations". Adaptive traits do not merely provide a benefit or perform a function; they are selected for their functionality (Sober 1984), they persist because of the benefit they provide (Bedau 1992, Bedau and Packard 1992). Ultimately, the functionality of adaptive traits promotes survival, reproduction, and, more generally, flourishing of the individuals. So, if the process of adaptation is the creation of adaptive traits, then open-ended or supple adaptation consists of the continual evolution of new adaptive traits.

But how can supple adaptation of this kind be measured in actual systems, especially if we are unsure which traits in the system are adaptations at all, much less adaptations for some specific kind of functionality? The difficulty&endash;some would say impossibility&endash;of answering this question has been stressed in a classic paper by Gould and Lewontin (1979). I have a proposal, developed in collaboration with Norman Packard (Bedau and Packard 1992), for how to answer this question. Our proposal rests fundamentally on the idea that we can detect whether a trait is an adaptation by measuring the extent to which the trait persists in the face of selection pressures. Roughly, the argument goes as follows. Whenever a trait is "used" or expressed, selection has an opportunity to provide feedback about the trait's adaptive value, its costs and benefits. If a trait persists and spreads through a population when it is repeatedly used, and especially if the trait accumulates significantly more use than one would expect to see if it had no adaptive value, then we have positive evidence that the trait is persisting because of its adaptive value. This is precisely what it is to be an adaptation. Now, since open-ended or supple adaptation is simply the continual evolution of new adaptive traits, we can exploit our evidence for adaptive traits into a straightforward sign of supple adaptation, as follows: If we continually see (on a relatively long time scale) new clusters of traits that are persistently used (on a relatively short time scale) significantly more than would be expected in the absence of adaptation, then we have positive evidence for the occurrence of the process of supple adaptation.

This line of argument needs certain qualifications and amendments. In certain special circumstances, a nonadaptive or even maladaptive trait can accumulate significantly more usage than one would expect of a "generic" nonadaptive trait. For example, the trait might be genetically linked with another trait that is adaptive. When a nonadaptive trait "hitchhikes" in this way on an adaptive trait, then the usage of the pair of traits will exceed that expected of a pair of nonadaptive traits. Thus, although we would need additional information to discern which trait is an adaptation, we will still have positive evidence that at least one of them is. So we can still interpret the continual occurrence of new traits with significantly elevated usage counts as the sign of supple adaptation. (Below I discuss another important qualification of this argument.)

To implement this proposal, one must collect statistics about traits' "usage", i.e., the extent to which their adaptive value is tested by selection. It is simple to collect and display usage data for the traits in many systems. Here I will provide just a brief summary and two quick illustrations of the methods involved; for more details, see Bedau and Packard (1992) and Bedau (1995). Although I will not develop the full generality of the method, it is worth pointing out that the method can measure the adaptive value of "traits" at a variety of levels, including individual genes, clusters of genes, traits or capacities of individual organisms, or the full sets of features that define genotypes. In general, to reflect the adaptive dynamics in a population of items i, the first step is to create usage counters for the items. A usage counter u(i,t) is a bookkeeping device which keeps track of the extent to which item i has been "used" and so tested by selection during i's entire history through the course of evolution up to time t. With the help of such usage counters, one collects the usage of all the items in the whole system into one global usage distribution, U(t,u), where the value at the point (t,u) in the distribution U(t,u) is defined as the number (or proportion) of usage counters that are equal to u at time t, u(i,t) = u. This usage distribution, U(t,u), is a quantitative summary of a system's adaptive dynamics.

The details of usage bookkeeping vary from case to case, depending on the kind of items involved and the kind of system they are in. For example, if i is a particular genetically-controlled behavioral trait an organism, then u(i,t) is the "usage" that behavior i has accumulated up to time t. Since the usefulness of a behavioral trait i is tested by natural selection whenever that behavior is exhibited, it would be appropriate in this case to increment u(i,t) by the number of times trait i occurs at time t; Bedau and Packard (1992) illustrate one way to measure usage of such behaviors. Or, to shift to a higher level of analysis, if i is a genotype present in a population, then u(i,t) is the usage of a genotype i at time t. In this case, since the viability of a genotype i is tested by natural selection to the extent to which i exists in the population (or, as we might say, i is "used"), it would be appropriate to increment u(i,t) by i's concentration in the population at t. In this case, then, a genotype's usage is just its integrated concentration, that is, the sum of its concentration throughout its history in the population.

The simplest way to observe a usage distribution U(t,u) is by graphing the value of U(t,u) as a function of time t and usage u, such as in Figures 1 and 2. The most prevalent structure seen in usage distributions is "waves": peaks of high usage that move across the U(t,u) plane. In general, each significant usage wave is the sign of an adaptation. Various characteristic kinds of wave phenomenology highlight different aspects of the dynamics of adaptation. It is easy to collect usage data in artificial life models, so simulations can easily generate lots of raw data for analysis. I have measured usage waves in a half-dozen computer models, including a model of evolving sensorimotor functionality (Bedau and Packard 1992, Bedau 1995), tierra (Ray 1992), avida (Adami and Brown 1994), echo (Holland 1992), the iterated prisoner's dilemma with noise (Lindgren 1992), and the bar problem (Arthur 1994). The general conclusion repeatedly confirmed in this work is that usage waves vividly depict the dynamics of supple adaptation. This makes usage waves a convenient empirical yardstick for comparing supple adaptation in different systems, including natural systems.

As an illustration of this technique, consider the genotype usage waves in one model&endash;Tom Ray's tierra (Ray 1992)&endash;shown in Figure 1. Tierra consists of a population of self-replicating machine language programs that "reside" in computer memory consuming the "resource" CPU time. A tierrian "genotype" consists of a specific type of string of self-replicating machine code, and each tierrian "creature" is a token of a tierrian genotype. A simulation starts when the memory is inoculated with a single self-replicating program, the "ancestor", and then left to run on its own. At first the ancestor and its off-spring repeatedly replicate until the available memory space is teeming with creatures which all share the same ancestral genotype. However, since any given machine language creature eventually dies, and since errors (mutations) sometimes occur when a creature replicates, the population of tierra creatures evolves. Over time the "ecology" of tierrian genotypes becomes remarkably diverse, with the appearance of fitter and fitter genotypes, parasites, and hyper-parasites, among other things.

To collect genotype usage statistics in tierra, we implement "usage" counters for each genotype, as outlined above, so that the value of the usage counter u(i,t) for genotype i at time t is simply the sum (more precisely, the integral) of i's concentration from its original appearance in the population up to time t. Then, the global genotype usage distribution, U(t,u), of a given tierrian simulation accumulates information about the temporal changes in all the genotype usage counters. For simplicity, we can let the value of the distribution U(t,u) at the point (t,u) be equal to zero unless there is at least one genotype i such that u(i,t) = u, in which case U(t,u) is set to one. Under this scheme, as a genotype enters the population and become more prevalent it will trace a line (a genotype usage "wave") moving up and to the right in the U(t,u) graph. This line's slope will change as the genotype's concentration in the population changes, and when the genotype goes extinct the line will end.

With this background we can begin to appreciate the significance of the genotype usage waves the model generates. Even a cursory study of a tierrian U(t,u) graph, such as Figure 1, reveals a host of significant adaptive events.

• The ancestral creature is clearly visible as the first wave in Figure 1, coming out of the origin of the usage distribution.

• The ancestral usage wave ends when the genotype goes extinct around time or "update" 500. This coincides with the start of a new usage wave, the second one in the run. Microanalysis reveals that the second genotype is fitter than the ancestor because it is shorter and so reproduces faster. The third main event is the biggest usage wave so far, which turns out to be a still shorter genotype.

• At around update 3500 a new kind of wave dynamics starts. Instead of single waves we see threads consisting of interwoven waves. The bottom of Figure 1 is a blowup of U(t,u) during this epoch. The separate waves in each thread are distinct genotypes, but they are acting in concert with all the other waves in the same thread. Microanalysis reveals two interesting facts: time step 3500 corresponds to first introduction of a significant population of parasites, and the different genotypes in a single thread are neutral variants which differ only in unexpressed machine language instructions.

• At around 15000 the biggest wave we see starts. This wave quickly dominates the entire usage distribution because this new genotype dominates the rest of the run. (The graph in Figure 1 clips off the top of this wave after update 20000.)

Without deciphering any more structure in these usage waves, we can already see how they sensitively reflect the significant adaptive events and co-evolving selection pressures that emerged in this tierra simulation.

There is an interesting and important subtlety in this interpretation of usage waves, though. Not all usage waves reflect significant adaptations; even deleterious items can be expected to accumulate some usage. In fact, it is only because a trait has been used that selection has an opportunity to evaluate the trait's adaptive value. So, before we are entitled to interpret usage waves as a sign of an item's proven adaptive value, we need to discount that quantity of usage that could be expected if the item were nonadaptive. One can crudely address this concern by truncating usage below a certain threshold; in fact, this was done (at a rather arbitrarily chosen threshold) with the tierrian genotype usage shown in Figure 1. A full discussion of techniques for the expected usage for nonadaptive items is beyond the scope of this paper; Bedau and Packard (1992) and Bedau (1995) discuss this issue at greater length. Still, the idea behind these techniques can be vividly suggested by contrasting a certain kinds of usage distributions. For variety, and also to illustrate how usage statistics can be gathered in adaptive contexts which do not involve evolution, I will show usage distributions from a different computer model&endash;the bar problem, devised by Brian Arthur (1994) to illustrate a certain form of inductive reasoning.

The model for the bar problem contains a population of agents who decide each night whether to go to a certain bar. The agents' decisions are sensitive to the facts that, if the number of patrons at the bar is below a certain threshold then everyone there has a great time, but if the population exceeds the threshold the bar becomes unpleasantly overcrowded. To make a rational decision about whether to go to the bar on a given night, the agents evaluate and employ various "predictors". These predictors are hypotheses or heuristics about how to tell how many people will be at the bar on a given night; these hypotheses often use the population in the bar's recent history to predict the next night's population. For example, one predictor might advise that the bar population on a given night will equal the average of the population over the last five nights; another might advise that it will equal the "inverse" (total number of agents less the population at the bar) of the last night's population; etc. Each agent considers a variety of predictors; in general, the sets of predictors considered by different agents overlap but only partially. To decide what to do on a given night, each agent evaluates all of his predictors, selects the one that has most accurately accurately predicted the bar population recently, and then follows that predictor's advice.

The agents in the bar problem are considered to be "flourishing" to the extent that as many of them as possible have as much fun as possible, i.e., to the extent that each night the bar is full but not overcrowded. The process by which they choose predictors does not involve natural selection (at least not in the simple version of the model described here). Predictors do not reproduce or die; they exist forever. What changes, though, is the extent to which a given predictor is used by the members of the population. Thus, in the context of this model, a predictor i is adaptive at a given time to the extent that i's recent predictive success is causing agents to use i to govern their behavior and thereby flourish.

When the bar problem is simulated, one discovers that the bar population tends to fluctuate around the "overcrowding" threshold with a mean slightly below it (Arthur 1994). This global dynamic equilibrium in the population's behavior is achieved even though the agents as a group cannot all follow the same predictor and thereby synchronize their behavior. (If the agents were to all followed the same predictor, they would all do the same thing and thus make the bar either packed or empty; neither leads to any fun.) Thus, the global equilibrium in the bar census emerges even though the "ecology" of predictors actually used by the agents is continually shifting.

The process of agents adapting their behavior in the light of recent history can be visualized with a usage distribution. The basic idea is to collect data on the agents' use of the different predictors. A predictor i is "used" whenever an agent follows i's advice because of i's recent predictive success, so it is reasonable to increment u(i,t) by the number of agents that use i at t. Thus, u(i,t) records the total number of times that predictor i has been used from the beginning of the run up to time t. For simplicity, we can let U(t,u) be equal to zero unless there is at least one predictor i such that u(i,t) = u, in which case we let U(t,u) be one. This predictor usage distribution U(t,u) will show "waves" which reflect the changing frequency of predictor use.

Now, if want to determine which aspects of the predictor usage distribution U(t,u) reflect predictors' adaptive value, we need some way to identify those aspects of the distribution that are not due to adaptation. To depict this, we can first model a parallel random process, that is, a situation with exactly the same structure as the bar model except that each agent always selects which predictor to follow simply by choosing one at random (with equal probability from all of his predictors). Figure 2 shows the usage waves in a pair of bar problem simulations. All of the model parameters in the two simulations are the same, except that, while agents in one (on the right) select predictors on the basis of their predictive success&endash;this is the normal bar model&endash;in the other (on the left) agents select predictors at random. In effect, the model on the left parallels the bar model in all particulars except that the behavior is generated randomly. In this latter "null case" all adaptation is blocked; agents cannot rationally govern their behavior.

When we compare the two distributions in Figure 2, we see many differences. For example, we see that in the random bar model (left) all predictors are used with roughly equal frequency (depending on how equally they are distributed among the agents) and so cause roughly straight waves of roughly equal slope. In the normal bar model, when the agents' choice of predictor can adapt to recent history (right), the wave dynamics are quite different. A few perennially good predictors produce waves with unusually steep slope, but most usage waves shift between periods of steep and flat slope, often roughly in concert with other waves. This loose but coordinated zig-zagging in groups of waves is a sign of a metastable dynamic in an ecology of predictors&endash;clear evidence of the unpredictably co-evolving selection criteria that drives supple adaptation. The distinctive pattern in the distribution from the normal bar model (right) is an empirical picture of what supple adaptation can be like in the bar problem. Even without any more detailed analysis of the usage waves in Figure 2, we can sense how the difference in this pair of distributions measures the degree of adaptation in the right-hand distribution. The differences between these distributions can be characterized quantitatively (e.g., with the chi squared statistic), but it is also obvious to the eye.

The method of using usage waves to highlight supple adaptation implies many testable predictions about empirical data, including those from natural systems. The underlying form of these predictions is that usage distributions will highlight a system's adaptive dynamics. Since the details of the specific mechanisms affecting adaptation generally differ in different systems, so will much of the adaptive significance evident in usage diagrams from different systems. In tierra, for example, what it takes for a genotype to flourish changes if dramatically shorter or parasitic genotypes arise. By contrast, in Arthur's bar model, the adaptive value of a predictor changes as its recent predictive success waxes and wanes. As the salient qualitative structures in Figures 1 and 2 show, usage distributions in tierra and the bar model highlight these different kinds of change in selection criteria. Still, there is rather broad applicability of some general predictions concerning the correlation between certain kinds of usage waves and certain kinds of changes in the adaptive dynamics (extinction, invasion, predator-prey interaction, cooperation, etc.). For example, the explication of supple adaptation with usage distributions predicts that the fossil record will show dramatic bursts of new usage waves only during periods of major adaptive innovation, and an analogous prediction would apply to major adaptive innovations in a wide variety of other kinds of systems. In general, usage statistics depict supple adaptation in a manner that invites empirical test.

Usage distributions can be used to define macroproperties that measure various aspects of the dynamics of supple adaptation. For example, one can quantify the rate at which new adaptations (usage waves) are entering the system and persisting (Bedau and Packard 1992), i.e., the extent to which the system is exhibiting supple adaptation; call this rate a system's vitality, V(t). Intuitively, a system's level of vitality V(t) reflects the extent to which new significant adaptations are arising and persisting. So, if we view life as supple adaptation, we can use a system's vitality V(t) to define the degree to which it is living or involves life. By this sort of means, a system's usage distribution U(t,u) and vitality V(t) could figure centrally in the explanation of the system's supple adaptation, and perhaps even the explanation of the extent to which the system involves life.

6. Conclusions.

I suggest that we can explain and unify the diversity of life by means of the concept of supple adaptation. Furthermore, I propose that we explicate the concept of supple adaptation by means of usage statistics, as outlined above. If I am right that supple adaptation can be measured through usage statistics, then we have in hand a feasible and empirical, objective and public, repeatable and refinable method for detecting the extent to which systems exhibit life.

Even if we should not yet make final judgments, much can be said in favor of this view of life. For one thing, it brings explanatory order to lists of the central hallmarks of living phenomena. For another, it brings a rich set of resources to bear on the controversial questions about life. For example, if the essential feature behind life is supple adaptation, then we can attribute life's borderline cases to the indefinitely varied array of more or less supple forms that adaptation can take. In addition, even if we don't try to settle these controversies here, we can appreciate how much more tractable are the puzzles about the different senses of life at different levels of biological organization, about the relative roles of matter and form, and even about the connection between life and mind.

The concept of supple adaptation might not map neatly onto our everyday concept of a living organism. Indeed, there might well be no scientifically and philosophically justifiable equivalent for that role. Still, the concept of life as supple adaptation, when explicated by means of usage statistics, provides a non-circular and non-arbitrary account of the essential properties underlying all forms of life. Analysis of a system's usage statistics requires no prior knowledge of whether any form of life exists at any level in the system. And since a proliferation of significant usage waves shows that clusters of traits are persisting because of their adaptive significance, usage waves allow us to see clearly and measure precisely the supple adaptation that seems to define life at its most general.

The significance of these results extends beyond our understanding of life and suggests a more sweeping conclusion about how practices like the pursuit of artificial life can&endash;and, in time, will&endash;change the face of philosophy. Artificial life and philosophy make natural partners, with related goals and methods. Eschewing the accidental and the contingent, both are concerned with maximal generality and essence. In addition, artificial life's emergent thought experiments are a natural extension of traditional philosophical methods. Artificial life will affect philosophy in a variety of ways: by facilitating progress on deep and traditional philosophical questions, some of which, like the nature of life, are presently being ignored; by focussing attention to important new questions, like the nature of supple adaptation; and by bringing a new level of clarity and precision to these issues, with empirical and quantitative techniques like usage statistics. These changes will not supplant philosophy by science but will allow the two disciplines to combine their resources in the pursuit of fundamental questions, like the nature of life.

Acknowledgements. Thanks to Hugo Bedau, Maggie Boden, Clif Bowen, Mark Hinchliff, Norman Packard, David Reeve, Teresa Robertson, and anonymous reviewers for discussion and comments on the manuscript; thanks also to Titus Brown for Figure 1 and Robert Seymour for Figure 2.

Adami, C., and Brown, C. T. (1994), 'Evolutionary Learning in the 2D Artificial Life System "Avida"', in Brooks and Maes (1994), 377-381.

Arthur, W. B., (1994), 'Inductive Reasoning and Bounded Rationality', American Economic Review 84: 406-411.

Bagley, R., and Farmer, J. D. (1992), 'Spontaneous Emergence of a Metabolism', in Langton et al. (1992), 93-140.

Bagley, R., Farmer, J. D., and Fontana, W. (1992), 'Evolution of a Metabolism', in Langton et al. (1992), 141-158.

Banzhaf, W., and Eeckman, F. (1995) (eds.), Evolution and Biocomputation&endash;Computational Models of Evolution (Berlin).

Bedau, M. A. (1991), 'Can Biological Teleology Be Naturalized?', Journal of Philosophy 88: 647-655.

Bedau, M. A. (1992), 'Philosophical Aspects of Artificial Life', in Varela and Bourgine (1992), 494-503.

Bedau, M. A. (1995), 'Three Illustrations of Artificial Life's Working Hypothesis', in Banzhaf and Eeckman (1995), 53-68.

Bedau, M. A. (forthcoming), 'Weak Emergence', in Tomberlin (forthcoming).

Bedau, M. A., and Bahm, A. (1994), 'Bifurcation Structure in Diversity Dynamics', in Brooks and Maes (1994), 258-268.

Bedau, M. A., and Packard, N. (1992), 'Measurement of Evolutionary Activity, Teleology, and Life', in C. Langton et al. (1992), 431-461.

Bedau, M. A., and Seymour, R. (1994), 'Adaptation of Mutation Rates in a Simple Model of Evolution', in Stonier and Yu (1994), 37-44.

Bourgine, P. and Varela, F. (1992), 'Towards a Practice of Autonomous Systems', in Varela and Bourgine (1992), pp. xi-xvii.

Brooks, R. and P. Maes (1994) (eds.), Artificial Life IV (Cambridge, Mass).

Cairns-Smith, A. G. (1985), Seven Clues to the Origin of Life (Cambridge).

Crick, F. (1981), Life Itself: Its Origin and Nature (New York).

Dennett, D. (1994), 'Artificial Life as Philosophy', Artificial Life 1: 291-292.

Dennett, D. (1995), Darwin's Dangerous Idea: Evolution and the Meanings of Life (New York).

Emmeche, C. (1992), 'Life as an Abstract Phenomenon: Is Artificial Life Possible?', in Varela and Bourgine (1992), 466-474.

Emmeche, C. (1994), The Garden in the Machine: The Emerging Science of Artificial Life (Princeton, N.J.).

Farmer, D, and A. Belin (1992), 'Artificial Life: The Coming Evolution', in Langton et al. (1992), 815-840.

Farmer, J. D., Lapedes, A., Packard, N., and Wendroff, B. (1986) (eds.), Evolution, Games, and Learning: Models for Adaptation for Machines and Nature (Amsterdam).

Gould, S. J. (1989), Wonderful Life (New York).

Gould, S. J., and Lewontin, R. C. (1979), 'The Spandrels of San Marco and the Panglossian Paradigm: A Critique of the Adaptationist Programme', Proceedings of the Royal Society B 205, 581-598.

Haldane, J. B. S. (1937), 'What is Life?', in his Adventures of a Biologist (New York), 49-64.

Holland, J. H. (1992), Adaptation in Natural and Artificial Systems (Cambridge, Mass.).

Kaneko, K. (1993/1994), 'Chaos as a Source of Complexity and Diversity in Evolution', Artificial Life 1, 163-177.

Küppers, B.-O. (1985), Molecular Theory of Evolution: Outline of a Physico-Chemical Theory of the Origin of Life (Berlin).

Langton, C. (1989a), 'Artificial Life', in C. Langton (1989b), 1-47.

Langton, C. (1989b) (ed.), Artificial Life, Addison-Wesley.

Langton, C. (1992), 'Life at the Edge of Chaos', in Langton et al. (1992), 41-91.

Langton, C., Taylor, C., Farmer, J. D., and Rasmussen, S. (1992) (eds.), Artificial Life II (Reading, Calif.).

Lindgren, K. (1992), 'Evolutionary Phenomena in Simple Dynamics', in Langton et al. (1992), 295-312.

Lovelock, J. (1988), The Ages of Gaia (New York).

Maturana, H., and Varela, F. (1973), Autopoiesis: The Organization of the Living (Dordrecht).

Maynard Smith, J. (1975), The Theory of Evolution, 3rd edition (New York).

Maynard Smith, J. (1986), The Problems of Biology (New York).

Mayr, E. (1982), The Growth of Biological Thought (Cambridge, Mass.).

Mayr, E. (1988a), Towards a New Philosophy of Biology: Observations of an Evolutionist (Cambridge, Mass.).

Mayr, E. (1988b), 'Is Biology an Autonomous Science?', in Mayr (1988a), 8-23.

Mitchell, M., and Forrest, S. (1994), 'Genetic Algorithms and Artificial Life', Artificial Life 1: 267-289.

Monod, J. (1971), Chance and Necessity (New York).

Packard, N. (1989), 'Intrinsic Adaptation in a Simple Model for Evolution', in Langton (1989), 141-155.

Poundstone, W. (1985), The Recursive Universe (Chicago).

Ray, T. (1992), 'An Approach to the Synthesis of Life', in Langton et al. (1992), 371-408. Reprinted as Chapter 3 of this volume.

Rosen, R (1991), Life Itself: A Comprehensive Inquiry into the Nature, Origin, and Fabrication of Life (New York).

Ruse, M. (1989) 'Introduction', in his Philosophy of Biology (New York).

Schrödinger, E. ( 1969), What is Life? (Cambridge).

Sober, E. (1984), The Nature of Selection: Evolutionary Theory in Philosophical Focus (Cambridge, Mass.).

Stonier, R. J., and Yu, X. H. (1994), Complex Systems: Mechanisms of Adaptation (Amsterdam).

Taylor, C. (1992) '"Fleshing Out" Artificial Life II', in Langton et al. (1992), 25-38.

Tomberlin, J. (forthcoming) (ed.), Philosophical Perspectives, Metaphysics, Vols. 10 and 11 (Altascadero, Calif.).

Varela, F., and Bourgine, P. (1992) (eds.), Towards a Practice of Autonomous Systems (Cambridge, Mass.). 

Genotype usage waves from the tierra model (Ray 1992). Above: a graph of the genotype usage distribution, U(t,u), cropped at time 30000 and usage 200. Below: a blowup of the epoch (roughly 3500 - 20000) that shows threads of interwoven waves. Genotype usage is defined as the integrated concentration of the genotype, i.e., the sum of its concentration through history. Time is measured in model "updates". Only genotypes with concentrations above a certain (low) threshold are shown.

Waves in usage distributions from the bar problem (Arthur 1994), illustrating how supple adaptation differs from a parallel random process. Right: agents follow their most accurate predictor of recent history. Left: a parallel process except that agents randomly choose from among their predictors. All other model parameters for the two distributions are exactly the same. The axes in these usage graphs (depicted only implicitly) are just like those in Figure 1; time increases horizontally to the right (100 time steps are shown), and usage increases vertically (but is cropped above 500).