Foreword

Though there were precursors, it is reasonable to date the conception of modern, computer-based studies of artificial life from studies of self-reproduction by von Neumann and Ulam (von Neumann, 1966). More than a quarter of a century later, Chris Langton was both the midwife and the person who named the offspring (Langton, 1989). From the outset, reproduction and, implicitly, adaptation – the basis of Darwin’s great theory – were features of artificial life. This book, published half a century after von Neumann’s unfinished manuscript became available, underlines the centrality of these ideas. Though the field is still far from its goals, the excitement remains. Whatever the long-term outcome, we can only agree with Eddington’s statement, ‘The contemplation in natural science of a wider domain than the actual leads to a far better understanding of the actual’ (Levy, 1992).

In both life and artificial life, perpetual novelty is a feature that continually comes to the fore. Of course, perpetual novelty in this context means much more than simple random variation. Still, it is not as mysterious or as difficult to attain as it might seem. Fewer than a dozen rules suffice to define the game of chess, but no two games of chess are alike (barring deliberate repeats) and, after centuries of study, we still discover new principles for playing the game well. Much the same can be said of Euclidean geometry, with new, non-trivial theorems being discovered after more than two millennia of close study. Is this kind of perpetual novelty related to that of life and artificial life?

My answer is yes and, moreover, I think the relation is close. Rules, axioms and computer instructions serve as generators that define a set of possible configurations, be they legal arrangements of pieces on a game board, strings of symbols that define theorems, or programs that direct a general-purpose computer. Similarly, molecular genetics makes it clear that interacting genes, defined by the 4-nucleotide DNA code, act much as a program with many conditionals, organizing the substrate into a network of reactions that maintains and defines a biological cell. Of course, there are many levels of organization in a cell and several different dynamics. That is what makes genomics and proteomics so complex, but most scientists believe that the perpetual novelty we observe in biology can be explained in terms of these generators. And, however complex the models reported here, they are all generated by programs based upon a small ‘alphabet’ of computer instructions.

What should we expect, then, of these computer-based models? Most computer-based models originate in an attempt to answer some question or cluster of questions. The art of model construction is to choose the mechanisms or rules defining the model – its generators – so that the configurations they define map clearly onto the area defined by the questions. In the case of living systems, the questions are so difficult that it is often a major triumph to discover any set of mechanisms sufficient to generate the phenomenon of interest. Von Neumann’s paper is a case in point. Until that paper appeared, most philosophers held that self-reproduction was a defining characteristic of life, using a recursus ad infinitum argument that a machine could not reproduce itself: the machine would have to have a plan of its organization, but that plan would have to include the plan itself, and so on. Von Neumann’s model offered an existence proof of a self-reproducing machine, thereby neatly collapsing sophisticated philosophical arguments to the contrary.

Much of the most productive work in artificial life has this proof-of-principle character. Without attempting to match living systems in detail, such models show that some set of mechanisms or rules is sufficient to generate the phenomena of interest. Because it is difficult to find any set of sufficient mechanisms for most lifelike phenomena, this is already a useful step. It is not ‘just’ a theoretical step either. Theories tell us where to look and so it is with these existence proof models. We can use experimental methods to see whether such mechanisms have counterparts or variants in real systems. And even if the living system does not have counterparts, the differences may well suggest a new set of mechanisms that do have counterparts.

Computer-based models are important in this endeavour because of the generated nature of most living phenomena. A fertilized egg provides a wonderful example. The rules and mechanisms embedded in that single cell can direct the generation of a complex metazoan, such as a primate consisting of billions of cells. A great variety of organisms are generated in this fashion, ranging from sponges to elephants. It is not easy to build analytic, equation-based models to describe this procedure. Even when it can be done, say with finite difference equations, the usual tools for analysing equation-based models, such as determination of fixed points or statistical analyses offer few insights.

Again, the game of chess offers a simple illustration of the difficulty of such analyses: a statistical analysis of the moves in a set of games will give few, if any, clues about the network of conditional decisions that defines a good strategy. Moreover, the fixed point of the game, the minimax value, is unknown, and of itself would tell us little about good strategies. Another metaphor strengthens the point: consider a porous meteorite plunged into a bath of liquid under high pressure. The fixed point in this case is a uniform distribution of the fluid throughout the cavities and pores of the meteorite, but the process of getting to this uniform distribution – with the near-surface, porous cavities being penetrated first, and so on – is a different matter entirely. The fixed point tells little about the trajectory of changes leading to it. In the case of living systems, the endpoint is death, which tells us little about the process of living. We need an entirely different kind of analysis to determine the characteristics of living systems.

What do computer-based models offer that is different from analytic equation-based models? They are executable. That is, we can observe the generated trajectory as it unfolds. Of equal importance, we can restart the model from selected, well-defined starting points, observing the changes imposed by modifications of the generators and rules of interaction. Though we lose the generality of good analytic models, such as the Lotka-Volterra equations, we do not lose rigour, and we gain in return great amounts of information about the trajectory. We also gain new insights into macro-phenomena such as robustness, speciation and the origins of autocatalytic reactions – phenomena that are critical to an understanding of living systems.

Most importantly, computer-based models, when successful, clearly demonstrate the emergence of macro-phenomena from micro-phenomena. Here we come to a question that is currently a subject of much debate in science and the philosophy of science: can we explain living phenomena with a reductionist approach? This question actually requires answers to two closely related questions:

(i) Can observable macro-phenomena be produced by a relatively simple set of generators and interactions?

(ii) Can we provide well-defined ways for macro-phenomena to influence the interactions of the generators?

There are agent-based models that give a proof-of-principle that both questions can be answered in the affirmative (see, for example, Arthur et al., 1997). Indeed, we do not yet have any clear examples of questions for which reduction must fail. Under the circumstances, it seems reasonable to me to press as hard as we can with this kind of reduction until we are stopped in our tracks by formidable barriers. Even then, given the great successes of reduction, it would seem reasonable to retain some optimism that new variants of the tools of reduction will find ways through such barriers. In any case, in the study of life and artificial life, we are far from ‘end of science’, even when it is interpreted as coextensive with reduction.

In model-based studies, it is important to avoid premature criticism centred on ‘what has been left out’. The art of model building is similar to the art of cartooning: the object is to select, and even exaggerate, salient points in order to concentrate on the questions of interest. The well-known Crick-Watson-Franklin story of modelling DNA’s structure makes the point (Maddox, 2002). And, as Ian Stewart says in his chapter in this volume, ‘It is not useful to criticize these models on the grounds of what they leave out: the test is what they predict, and how well it fits the appropriate aspects of reality, with what they leave in’. Only in this way can we begin to build the grand interaction between theory and experiment that will let us tame the floods of data being produced by the new tools of biology.

There are many fundamental questions in the study of life and artificial life that are still open. Several of them are the subject of papers in this volume. I will look at just three examples:

(i) Ian Stewart makes a strong case for the relevance of a theory of pattern. From the days of D’Arcy Thompson (Thompson, 1961) and Weyl (1952) we have known the critical role of pattern in understanding complex systems. But, for...