Leonardo

Three Perspectives on Artificial Evolution

Participants in this interdisciplinary research area have explored artificial evolution according to various interests. Very roughly we could distinguish three main tendencies in the research:

• • In the first, evolutionary theory is used as a problem-solving strategy, convenient for certain purposes, but not requiring biological plausibility (because the procedure is not being applied to generating life or lifelike phenomena). As the main goal of this case is the efficiency of the method used to achieve practical goals, both theoretical and empirical knowledge can be violated (for example, working with Lamarckian evolution or very high mutation rates) in order to achieve certain goals, such as optimization.

• • A second tendency uses artificial models to study biological phenomena for which no adequate model or theory has been developed so far. The aim in this case is to develop modeling strategies to enhance our capabilities of explaining biological phenomena, thus generating theory or analytical tools based on artificial systems.

• • A third division is constituted by those artificial systems in which the frontiers or boundaries between the artificial and the biological are diffuse, and instances of alternative universes are constructed that, although not corresponding literally with the facts of real life, provide an intuitive grasp of lifelikeness.

These three categories correspond roughly to the production of three different things: tools, models and instantiations [9]. Tools are methodological innovations that may or may not apply to life, although inspired by it; whereas models and instantiations are distinguished by their respective relations to a real phenomenon to be explained or grasped. The term model is reserved for the case usually developed by the natural sciences: a way of representing certain relevant aspects of an empirical domain by means of data collected through measurements. In this case the natural systems are, so to say, the referents of the artificial models built to explain or emulate them. This reference relation is looser in instantiations, more related to the way models are conceived in formal sciences: as an instance or a system built after a set of axioms. Because they do not have a concrete empirical referent, their value resides in their being new creations of ontology that, although inspired by natural phenomena, provide as an end result something different from life on Earth, but with a strong lifelikeness.

The three kinds of artificial systems have different goals. Tools enhance our epistemological capacities in the sense that they provide information-processing ways that, although inspired by the biological theory, do not operate in a way transparent to the human mind. They are directed at fulfilling human or technological goals. Models may allow us to study the distance between realized and non-realized life, the conditions under which other possibilities for our life could have happened and counterfactuals based on different parameters. They may provide theoretical contributions if the data is applied to the appropriate experiments. Instantiations may be much more abstract than models and their only connection with reality depends on the way we apply the theory we have. Although it may be the case that instantiations will provide the most radical ways of exploring life-as-it-could-be, present-day work is strongly limited by its dependence on difficult-to-overcome theoretical assumptions.

Intuitive Notions of Evolution and Evolutionary Theory

Models and instantiations are, in principle, more likely to produce a notion of lifelike artificial evolution than will tools. At the moment, most of the research in artificial evolution is based either on an intuition or on a theory, or on both. The intuition is that evolution, whatever it is, must reveal itself as the growth of some-thing: complexity, adaptation, optimization or some other global property. Even if this view of evolution as growth tends to be rejected by the latest evolutionary theory (where the usual perspective is neutral with respect to this), it remains an important guiding notion for a science that has to provide artificial examples of something so poorly understood as evolution. Evolutionary biology considers that evolution takes place when there is a change in the gene frequencies of a population (the Hardy-Weinberg Law), change driven by different "evolutionary forces": natural selection, sexual selection, genetic drift, etc. This notion of evolution assumes a fairly neutral situation with respect to what is expected from evolution, whether or not there is an increase in complexity in evolution, or whether evolution has a direction. Most artificial evolution models are not that neutral.

The motivation of researchers trying to study or to implement artificial evolution might in fact preclude that neutrality. For instance, when von Neumann [10] studied the problem of artificial self-reproduction, he took into account the possibility of an increase in complexity of the reproduced system. In his view, the capacity of living systems to evolve derives from the possibility that they, unlike machines, have to produce other systems that are of equal or superior complexity; and many of the constraints he imposed upon the logic of self-reproduction, such as the necessity of a universal constructor, were required by this conception of open-ended evolution. Similarly, evolutionary computation, as a general problem-solving tool, assumes that this methodology will bring along some sort of "improvement" or optimization of the proposed solutions.

Thus, A-Life experimenters trying to develop artificial evolution face a dilemma: provide an external goal to their systems or have no natural "goal" at all. Many artificial systems are indeed supplied with an external goal; this is the case with most tools and also with exploring creativity that is to be evaluated by humans (an example could be art created through this procedure). However, the case of an artificial evolution that is to remain true to nature or lifelikeness is more complex, for, on the one hand, the evolving system supposedly has no intrinsic direction and, on the other, it is assumed that given the [End Page 276] appropriate evolutionary dynamics (and the appropriate encoding of the problem) any result could be achieved or is possible. The reason for the latter is that the kind of abstraction of natural evolution used by artificial evolution must be able to derive any interesting system or organization as its evolution under the appropriate variety and the selection pressure. This is why in general a procedure of artificial evolution, such as the genetic algorithm, gives the impression that anything is possible. Yet such a situation is equivalent to one in which nothing interesting happens. However, this is not the case in natural evolution; the reason is that its intrinsic trends have not yet been sufficiently investigated by artificial systems.

This general idea can be expressed in another way by saying that the evolution should be based on synchronically complex or organized systems. The study of complex systems presents a dichotomy of methodologies: the approach taken is either synchronic or diachronic. The synchronic one (sometimes also called vertical or emergent) studies the relation of components to the aggregates or totalities they form; the target is complexity or interesting behaviors at a global level, starting from simple components or interactions at a lower level. It is a bottom-up perspective, largely inspired by thermodynamics. Examples of this approach are auto-catalytic sets, swarm organizations, neural nets and embodied robotics. The diachronic approach (also called horizontal or transformational) studies how complex systems arise and develop in time, as a substitution of subsequent generations. The system is considered in terms of the appearance of novelty through time, and not in relation to levels, selection being the main mechanism or force acting on the system. Work in artificial evolution, such as the genetic algorithm, is an example of this diachronic approach.

In many cases artificial evolution seems to rely on intuitive notions of evolution, instead of trying to develop new, productive ideas to explore what evolutionary change could be like for artificial systems. In order to develop these novel procedures, we require an appropriate theory of organizations able to evolve, as opposed to random or structured bit strings.

Organization, Functionality and Design

The lack of a well-developed conception of diachronic embodiment is evident in the treatment of functionality. The functional analysis of complex systems is a hard issue for biology and A-Life that can be explained, even if in a sketchy way, by referring to the difference between two traditions of thought, the Kantian and the Darwinian, through the way each of them compares the living organism with a watch.

For the Darwinian tradition, this comparison poses the problem of the "argument from design" (developed by, among others, Aquinas, Hume and Paley). Paley wrote that if in crossing a heath one found a watch one would not think that (like a stone) it had lain there forever, but would infer that the watch must have had a maker: "Arrangement, disposition of parts, subserviency of means to an end, relation of instruments to a use, imply the presence of intelligence and mind" [11]. In the same way, the design of organisms brings us necessarily to accept the existence of a creator: "Every manifestation of design which existed in the watch, exists in the works of nature, with the difference on the side of nature of being greater and more, and that in a degree that exceeds computation" [12]. When the Darwinian tradition responds to this argument, a natural explanation of design is produced-the principle of natural selection-that requires no divine intervention, but it accepts as good an analogy between the watch and the organism.

Kant had already used the same comparison in the Critique of Judgment [13], but in a rather different way. Kant observes a fundamental difference between the watch and the organism: whereas the watch is formed by fixed components, fabricated beforehand and later assembled, in the organism all parts are formed so as to link with the others, and some parts produce the others. Kant accepts an internal teleology in the living system.

Many authors (for example, Mayr) have said that the later development of the theory of evolution corrects the Kantian summons to teleology and makes it possible to explain biological beings without it. Maybe this is true, but Kant points to a problem that the Darwinian tradition has not addressed: how the relation among the parts forms an organization. Actually, as in the case of the watch, for the Darwinian tradition the assimilation of watch and organism is not problematic, whereas in the Kantian one this distinction must be made and explained.

A consequence is that it is not the same to explain the function of components in relation to organization or, as is usual in evolutionary biology, to adaptation (either historical or a-historical). The two traditions approach the problem of organization through an analysis that derives the whole from the properties-decomposable or not-of the parts. However, it might be better to conceive a complementary "synthetic" (creative or productive) way of looking at this problem.

This is the reason why embodiment-both synchronic and diachronic-is such an important issue for A-Life. Current proponents of "embodied cognition" seek to understand how physical interactions contribute to the autonomous behavior of robots or artificial agents [14]. The suggestion is that when the physical properties of bodies (shape, orientation, degrees of freedom, etc.) are taken into account, a great deal of abstract information processing becomes superfluous, because the relevant behaviors emerge out of the physical interactions of the system with the environment. If more elements related to the body of these agents were considered-for example, metabolism-more radical changes would come about [15]. When valid explanations have to be offered mainly in terms of mental capacities, physical interactions are conceived as imperfections, as handicaps or constraints that limit information processing capabilities. Embodiment precludes the search for "clean" functions, with serious consequences for design. If evolution were a process of creation of functional-ities (that is, a process in which matter is shaped to produce-and reproduce-certain functions), in most cases these would be non-detachable properties of the physical substratum.

The problem of functions leads to that of design. Every theory of evolution aims to provide natural mechanisms for the origins and diversification of living forms, so that no designing agent has to be proposed. Evolution is intuitively different from design in its non-intervention. Evolutionary explanations are naturalistic because form or function emerge from material interactions. Naturalistic explanations of design are, nevertheless, complex, even if it is agreed that the process of evolution is not designed and, moreover, that it does not itself design. For machines, the case is easier, but the aim of A-Life is to produce lifelikeness. Let us turn for a moment to Polanyi's ideas about machines. He described them as working under two distinct principles: a higher one of design, and a lower one, harnessed by the former, of the basic [End Page 277] physical-chemical processes [16]. This account may hold for living beings as well, but their design is not easy to characterize, because both internal, organizational and external, selection-based factors take place. Then the attempt to substitute design for artificial evolution in the construction of complex machines, such as adaptive robots, requires one to start with no functions, to let them emerge.

Artificial Evolution and Natural Selection

In A-Life, artificial evolution has often been used to build artificial creatures with features adapted to their artificial environments by a procedure similar to evolution by natural selection (for example, the genetic algorithm). The methodologies developed with this inspiration start generating (at random or by craft) a population of bit strings that are interpreted as possible solutions for the problem; then the "fitness" of each of them is scored using an evaluation function. This makes possible a process of selection of individuals according to these scores (usually the best are used as the "parents" of the subsequent generation, but sometimes samples of medium or even substandard individuals are also preserved). Finally, individuals are modified using the genetic operators (mutation and/or recombination) to produce a new generation. This is the basic genetic algorithm; if one lets the system "evolve" in this way, the population usually converges into a situation of mutual environment/inhabitant adaptation. In the artificial worlds domain, this kind of artificial evolution dictates changes in certain features of simulated agents and the environment; it introduces a higher scale of temporal change than that of processes at the scale of the lifetime of individuals (for example, learning). In the most interesting cases, the interaction between both scales produces interesting phenomena. In the evolutionary robotics domain, this method has been conceived either as an alternative to the explicit design of the morphology and/or the cognitive systems of artificial creatures or as a design methodology (there is some ambiguity about which is the case).

In the first domain, the artificial world is usually a two-dimensional array of cells (but it can be n-dimensional, or sometimes, continuous) where a population of agents coexists with similar ones and other living (predators, parasites, etc.) and/or non-living systems (food, shelter, geographical barriers, etc.). Agents are typically represented by a control system for behavior specification (such as a neural network, a finite state machine, or an abstract grammar) and, very seldom, with a simulated physical body. They are endowed with a sensorimotor apparatus with which to perceive the relevant features of the world and a set of behaviors with which to act in it. The sensorimotor system is, in general, as complex as the agent's world: sometimes both are relatively realistic, while, in others, perception consists merely of detecting the state of the nearby cells. The metabolism of agents is usually represented as an energy storage system whose level determines whether they are in good shape to perform various actions (mate, escape, etc.) or close to death. Generational succession results in a modification of the control system (and/or the body) of the agents: as in evolution by natural selection, agents will be able to survive and reproduce according to their abilities in the world they inhabit.

Most of these models try to simulate the relevant features of evolutionary models (morphological, functional or behavioral traits that influence the fitness of individuals, size of populations, selection pressures, etc.) considered necessary for setting up experiments. These features are rather close to the phenomena studied by population genetics. The problem is that the change (that is, the evolution) likely to arise in simulations is, in general, already known from the theoretical results; also in general, the creativity or exploration of potential novelty expected from an evolutionary process is lacking.

These works do not offer new knowledge or intuitions about interesting ways in which evolution can influence complex (vertical) organizations. The main criticisms they have received refer to the problems of building simulations, because simulated agents do not have the physical organization necessary for the study of embodied cognition in the artificial domain. Critics of such work believe that research should be performed on physical realizations. In my opinion this criticism is correct, but insufficient, as it overlooks the problem of adequate artificial evolution, which is as important a factor as realism in the implementation of the dynamics of cognitive interaction.

In evolutionary robotics, the evaluation function depends on how well the robot behaves in the real world, instead of in a simulation [17]. This methodology combines simulations (populations of encoded neural nets evolve in computer simulations) and physical building (the performance of each individual neural net is transferred to a robot and evaluated according to its behavior). The rationale for adopting the methodology of artificial evolution tries to overcome problems associated to external functional decompositions of the system, in which the system is divided into parts according to the functions they perform in the whole system. Yet, it is not certain that these parts evolve independently: interactions among parts are many and complex, and sometimes they may be mediated by the environment. Thus, agents whose interaction with the environment is very sophisticated require a very complex sensorimotor system, difficult for hand design. That is why artificial evolution is used as the search method and also as the heuristic for handling the intractable complexity that up to certain levels arises in other hand-design incremental procedures [18,19]. This is also linked with the ideal goal of obtaining via evolution an incremental complexification of the behavioral capacities of the robots in an unforeseen way. Nevertheless, the separation of the diachronic and synchronic aspects of the process (performed respectively in simulation and realization), and the correspondent diachronic disembodiment, demand further investigation of artificial evolution.

If the artificial evolution procedure applied in these systems is analyzed, it is clear that eliminating the task of design is not that easy after all: many problems arise as a result of not intervening in the process from the outside. Issues such as how the genotypes of the population are encoded (and decoded), the relation between evolving populations and their environments and how the artificial models exhibit "natural" selection make apparent a fundamental lack of understanding of how a spontaneous, non-externally directed evolutionary process takes place.

The first problem, that of how to encode the evolving traits of agents in the genotype (only the encoded ones are affected by the genetic algorithm) is still poorly understood in most work. The genotype-phenotype relation tends to be very simple (in most cases direct) mapping and, normally, only a few genes are allowed to evolve, in fixed-length genotypes. Besides, genes evolve individually, there is no interaction between them in generation of the system. There is a growing awareness that the process of development has consequences for the way [End Page 278] genotypes evolve, and there have been interesting attempts to overcome these difficulties (for example, by using variable-length genotypes and, more important, developing more complex ways to represent the genotype-phenotype mapping, so that some aspects of development are taken into account). Nevertheless, important background questions are still underdeveloped. Von Neumann, in his early work on self-reproduction [20], wrote that only systems endowed with a self-description were capable of open-ended evolution. It was important to establish this threshold, but it still requires deeper understanding in the light of new research. In fact, the way von Neumann conceived the relation between the description and the constructed system is still a matter for study in the field, and various new approaches have recently been proposed [21].

The second problem, the relation between the agents and the environment as a common history, is also difficult to implement in such a way that conditions appear in which agents and environment inform one another and intervene in changing each other [22]. Simulated environments are too simple, fixed and external, and this makes it difficult to observe the action of agents to produce their own environments. Yet, even the real environments in which robots operate are usually externally controlled, not constructed by the creatures themselves.

Finally, it is very difficult to model the conditions in which selection can arise from the characteristics of the whole system-that is, to define the possible adaptive landscapes and intrinsic evaluation or fitness functions according to those characteristics. In fact, some models use straightforward (human) artificial selection, applying theoretical or aesthetic criteria, while others define conditions to be automatically evaluated, but with an absolute criterion, i.e. they pre-define a "perfection" to be achieved. Natural selection in artificial evolution is external and designed when a simple evaluation function scores individuals of the population. But it is not much more "natural" when creatures live, reproduce or die in this world according to their abilities, because the poor genotype-phenotype encoding and the simple agent-environment relation makes it rather obvious which of the parameters will define the best "fitness" in that world. An interesting solution for this has been sought in co-evolution, so that the fitness of a population is evaluated in relation to changing conditions [23,24].

For evolutionary biology, artificial selection is the process of change induced in a population by a human agent selecting something for some purpose. Natural selection, in turn, is blind, and even if some can consider it a process that produces adaptations, or even design, these are not foreseeable in advance: they depend on many changing conditions that shape the interaction between the evolving entity and the environment. In the case of artificial evolution, the situation is confusing, because the concept of natural selection is in some ways epistemologically very close to that of artificial selection. In fact, the metaphor of artificial selection in breeding was used by Darwin himself to propose natural selection as an explanatory mechanism for evolution. Perhaps the word "selection" cannot but suggest a selective agent, but it also might be that natural selection needs further examples (as suggested, for example, by Depew and Weber [25]).

Artificial Evolution's Contribution to Science

Some issues presented in the last section should serve as a qualification of the achievements of artificial methodologies whose main criterion is biological plausibility reduced to a standard textbook theory. This is an important problem for A-Life, because the phenomena under study are complex, and some degree of biological plausibility might seem appealing. However, an excessive attention to the biological might be a handicap, as in order to obtain insight into some phenomena under study, it is important to acknowledge the characteristics of the medium in which a model is produced. For example, embodiment, emergence, and complexity itself are difficult notions to understand, and copying biology is not necessarily the best way of clarifying them. In fact, the strategies of biology and cognitive science are often reductionist, and if research in the artificial domain is to overcome some of the problems encountered by these approaches, then other methods should be sought.

In biology there are theoretical problems in integrating the two apparently opposed functional and structural perspectives on organisms. The notion of self-organization is problematic for the standard theory of evolution. Self-organization, unlike the notion of natural selection, can be characterized as a systemic property, in the sense that the entity that organizes itself is composed of parts whose configuration and interaction determine the whole they form (which cannot be reduced to those parts). It is also a property that has to do with the production of spontaneous order in the system, which is a result of the dynamics of interaction among the components and has nothing to do with an external organizing agent or with external design. Finally, self-organization is a capacity that expands the explanatory domain of classical physics and chemistry. Hence, even if it is not a uniquely biological concept (it also appears in inanimate systems), self-organization provides a view of nature that suggests a continuity between the inanimate and the animate. These characteristics contrast with the atomistic, externalist and non-physicalist perspective of evolution by natural selection.

One possibility for artificial evolution, then, is to explore the integration of both perspectives. In this way, it would do more than provide computational versions of the analytical models already developed; it would contribute to the unsolved problems of the field. In fact, artificial models of biological systems have been proposed before; for example, auto-catalytic sets [26,27], Turing's reaction-diffusion model of morpho-genesis [28], and models of random Boolean automata have been relevant in reconsidering biological theory or at least in opening interesting debates within it. An application of the new ideas proposed in evolutionary biology to the domain of the artificial is important in order to expand the often too narrow application of evolutionary principles found in some artificial models of evolution, for example, in genetic algorithms.

Particularly relevant for the study of possible evolutions is the notion of developmental constraint as a "bias on the production of variant phenotypes or a limitation on phenotypic variability caused by the structure, characteristics, composition, or dynamics of the developmental system" [29]. The possible sources of those constraints are varied, but they all introduce intrinsic limits on the action of natural selection, either universal (for all life) or local (only in certain taxa). Developmental constraints, if they can be characterized, set limits to ideals of perfection: natural selection is not a mechanism of unlimited optimization; evolution does not produce "perfect design." Thus, developmental constraints are material limits to perfection.

Kauffman conceived the universal developmental constraints he studied as generic properties of organized matter [End Page 279] that do not depend upon natural selection, but which could influence the conditions in which it takes place [30]. These generic properties are based on the capacity for spontaneous order upon which natural selection acts. Self-organization is prior to the constitution of the system itself; it prepares the conditions in which natural selection can take place. Kauff-man has constructed mathematical models of genetic regulatory systems as logical networks of connections. By changing the connectivity parameters, Kauffman is able to study the conditions in which attractors and limit cycles appear. Intermediate systems, those found between order and chaos, have the most propitious landscape in which to evolve. This set of conditions defines a new null hypothesis to determine whether there is evolution; it may act as a substitute for the Hardy-Weinberg equation [31].

Yet evolution can also be understood in terms of stages as the model of generative entrenchment suggests [32]. This is based on a systemic standpoint: nature is divided into systems that are only partially decomposable and form several levels and unities. The idea is that natural systems are locked in stable ontogenetic paths that, once formed, cannot be reshaped again, except when there is a general reorganization of the major phylogenetic taxa. Development, then, makes evolution a quasi-irreversible process, in which each stage strongly determines subsequent evolution.

In addition, from a thermodynamic perspective, Brooks and Wiley present what they call a Unified Theory of Evolution [33]. A consideration of this may be useful in considering whether generic properties are best expressed in terms of dynamics or of thermodynamics. Dynamic descriptions are deterministic and reversible and require a detailed knowledge of initial conditions, while thermodynamic descriptions are stochastic and irreversible and require a selective description of initial data. Thermodynamic approaches to evolution may bring diachronic and synchronic perspectives together.

If artificial evolution is elaborated so that an evolving system is conceived as a complex material entity, change in embodied systems and the study of the generic properties of evolutionary change may be reunited. Many of the works in which artificial evolution is used to design robots or artificial autonomous agents recognize and try to advance the study of synchronic embodiment, but diachronic embodiment is not questioned. These works combine a self-organizing perspective of the structure and behavior of the agent with a limited selective perspective. Hence, elements that are disembodied in the diachronic sense are automatically introduced. For these reasons, a natural extension of the notion of embodiment is the development of new forms of artificial evolution to create a unified perspective.

The unified perspective contributes to embodiment because it can help to develop a different way of understanding functions as emergent and embodied constraints. Moreover, the philosophical discussion of whether they have ontological reality may be converted into other more concrete ones. For some theorists, functional or symbolic explanations only consider secondary properties, largely related to human observation and categorization, whereas others think that some natural systems (for example, the cell) produce their own functional or symbolic structures, that "stand for" longer reactions (or information pro-cessing) and can be transmitted as such in evolution [34]. The search for generic properties does not present an alternative between these positions. Underlying the property will be a certain specific relation between levels, which might be a non-detachable unit of evolution for that system. Thus, the investigation of functions as generic properties can integrate the diachronic and synchronic perspectives of embodiment, because it involves a further freeing from design (or a naturalistic understanding of it).

Counterfactual evolutionary processes are also interesting. The properties and questions proposed for such research have been of the type: Why are there only two sexes in most taxa? It is probably insufficient to investigate such a question by studying the various selection pressures that could have produced a variety of sexes. However, there might be insights to discover if the question of possible evolutions is posed in terms of self-organizing properties.

Conclusion: Evolution, Creativity and the Possible

Artificial evolution and research in autonomous systems inherit ideas from sources of distinctly differing epistemological styles. Artificial evolution has mainly followed standard evolutionary biology, while research is based on self-organization. An effort to expand the theoretical basis of artificial evolution should be made so as to achieve a better understanding of the evolution of organization. This would have interesting consequences for the way we understand evolutionary creativity and the domain of the possible opened by it. When creativity is conceived in a combinatorial way, almost anything is possible, but no interesting phenomena occur. The reason is that organizations are based on emergent properties that confer some kind of physical cohesion to the system, making it something different from a conjunction of parts or elements [35].

The different purposes driving artificial evolution seek creativity in different ways, but all of them-tools, models and instantiations-would benefit from a notion of constrained creativity that the evolution of organizations suggests. Non-intervening dynamical models are important in developing a notion of natural selection that is indeed natural, especially in the case of models of evolutionary phenomena. It is not so important to preclude conscious intervention in art, where perhaps the emphasis is placed more on using the evolutionary inspiration as a way to explore new forms of creativity, rather than on making the process biological. Yet, probably in both cases, biology and art, there is a similarity in the effort to understand the sources of creativity. In both biology and art, these sources are not completely free or combinatorial, but constrained in ways we would like to understand better. A constrained creativity not only limits variety but also enables novelty to appear in the system.