Short works

Books : reviews

Claus Emmeche.
The Garden in the Machine: the emerging science of artificial life.
Princeton UP. 1991

rating : 3.5 : worth reading
review : 3 April 2003

Emmeche, a theoretical biologist, takes a look at the new area of artificial life, and whether the "strong a-life" claims that it really is life, not just a simulation, are justified. His roving discussion covers von Neumann's self reproducing automata (reproduction), life at the edge of chaos, Lindenmeyer's L-systems (growth), Conway's Game of Life, Ray's Tierra (evolution), Reynold's boids (emergence), and more.

He has interesting things to say about the dynamics of emergent properties and self-organisation.

we can see that a whole range of evolutionary phenomena are internal, inherent properties of self-organized systems, and that natural selection need not lie behind all the structures that biologists observe in nature. ... selection can now also be understood as the means by which a dynamic system chooses its own trajectories or evolutionary paths.

The new construction of a complex system from identical rules and identical starting conditions can produce a somewhat identical result without explicitly coding the macroscopic construction.

The key phrase here is "identical starting conditions". When people get all computational about DNA, they go on about computer programs, but they rarely mention that other crucial aspect: the input. For life, that is the circumstances in which the DNA finds itself: a given, determined by the previous generation (the local environment of egg and womb, say), and by the external environment. As Jack Cohen so memorably puts it: "If you think it is just DNA, you don't worry about giving the mother thalidomide". The same DNA with different initial conditions (and a different environment) can result in different organisms:

Waddington himself made an attempt whereby ether treatment of fruit-fly larvae could induce the development of adult phenotypes that were exact copies of the kinds of bodily defective flies otherwise thought to develop only by mutations ... He further demonstrated that by continually changing environmental conditions, he could cause the genetic assimilation of these altered phenotypes; that is, the change would become stabilized via subsequent genetic changes. This was not accomplished by any Lamarckian inheritance but by a form of feedback between selection in the changed environment and embryonic development via other routes in the epigenetic landscape

And even if we do know both the program of life, and its inputs, there may not be any shortcuts to predicting the result (except by executing a totally faithful simulation -- not terribly feasible -- on a processor that runs faster than the real-time computation that life "performs").

For [computationally irreducible] complex systems there is no other way to predict their evolution than by tracking it. We must wait and see what happens. If this description also covers embryonic development, it means that the precise form of the final individual (the total phenotype) cannot be predicted, even with complete knowledge (which is itself purely hypothetical) of the total genotype and the epigenetic substrate.

Studying the physical limitations on computation entails acknowledging that computation, much as the measurement process in physics, is far from completely idealizable or meta-physical. Computational processes are subordinated to normal, natural laws. At the same time, these laws are described by mathematics and computations based on this mathematics. Hence, it is natural to conclude that the physical laws must have a form that permits them to be executed in time and space. The laws that describe our universe must be limited to algorithms that can in fact be executed using this universe's bricks and mortar. The universe does not permit us, for example, to build a computer with an infinitely large memory that could help us to separate pi from an arbitrary close-lying neighboring number.

He reiterates a useful distinction between two kinds of laws:

Perhaps we ought to revive Kant ... and the distinction he used in another context. This is the distinction between (a) systems that follow a rule, i.e., consult a representation, some symbols, and make computations based upon them ... ; versus (b) systems that operate in accordance with a rule, i.e., where the system behaves as if it had consulted a rule, but where the rule-like nature of the system ... is instead caused by purely natural lawlike causality. Even though my digestion is normally quite regular, my stomach is no computer, which for each new piece of meat searches its memory for a meat-digestion program. This does not prevent me from describing my digestive system as if it were a computer, however.

Artificial life follows laws (the computer program), whereas nature, including real life, acts in accordance with laws (the laws of physics, chemistry and biology). Emmeche says it is possible to describe a natural system as if it were a computer (Dennett might call this the "computational stance"). Yet what is to stop us equally well describing a computational system as if it were real (the "strong a-life" stance, perhaps?)

Emmeche is certainly no "carbon chauvinist" -- he admits the possibility of life based on other processes, but still only physical real-world processes, not virtual ones.

It is conceivable that life is not bound to a specific carbon-chain chemistry, but it is probable that the ways in which life will self-organize in alternative physical media-i.e., the concrete embodiment it achieves, the species that will appear, and the types of processes it contains-will depend on the specific physics that characterize the given medium. That is, the independence of life-forms from the medium (in the strong sense) is denied, but the possible realization of life in other physical media is not excluded.

It's the difficulty of simulating everything that Emmech appears to believe separates reality from artificial life.

Computer organisms do not die at all in the biological sense; they simply cease to exist. But real death is part of the living game. ... Life is also death, dissolution, dung, and earth.

In particular, it's the fact that the artificial life simulations are not "open".

of an artificial life program: we cannot speak of genuine open-ended evolution, for the system always seems to be caught in one attractor or another, where the condition is thereafter a very specific one.

Life is certainly a game, but it is not played according to fixed rules. The rules themselves can evolve, and during evolution old rules can break down.

genuine life is fundamentally different than computer life, whose environment is simulated. Computational forms of life do not stand in the same open, nonspecified relation to its environment because the environment, too, is of a purely computational nature: digital and determined.

"Openness" refers to the "gateway events" that open up new regions of phase space for the system to explore, so that it is not trapped in some static, bounded space. I have a lot of sympathy for the position that genuinely alive systems need to be open. However, just because existing simulations are not "open", does that mean that no simulation can be? Is openness a property only of the real world, or can it be a property of virtual worlds, too? I am not convinced this question is in any way settled. Emmech seems to be, however, and a late footnote, during the argument against strong a-life, has a revealing parenthesis:

Searle's Chinese Room argument against strong AI ... may not be logically sound in a strict sense (though still intuitively convincing), and his biologism ... with respect to mental states is debateable, I think that the argument against strong a-life can nevertheless be made over a similar, though not quite analogical, scheme.

So an argument against strong a-life can be made that is similar to Searle's logically unsound, and debateable, but "intuitively convincing" argument? Intuitively convincing to who? Several people, including Dennett and Hofstadter, might disagree about that convincingness.

I enjoyed the majority of this book, despite disagreeing with the main conclusion. But I have a problem with the last chapter, "Simulating Life: Postmodern Science". The book is translated from the Danish. Apart from a couple of bobbles -- Brooks' "fast, cheap, and out of control" is rendered as "quick, cheap, and out of control", and Conway's Game of Life is consistently referred to as "the Life game" -- the translation reads smoothly for the most part. This last chapter, however, might just as well still be in the original Danish for all I got from it. In previous chapters, my eyes passed over the text, and understanding appeared in my head; here, my eyes passed over the text. I don't know whether this is due to the translation, but I have a sneaking suspicion it is due to the content. I don't appear to have a Postmodern brain. (Judging from some things I have read about the subject, this is not a lack I lose any sleep over.)

All in all, there is a lot of interest here. Emmeche brings an thoughtful view to the a-life debate, and raises lots of intriguing issues. I am not convinced, however, that he has made his case against the strong a-life view.

Peter Bogh Andersen, Claus Emmeche, Niels Ole Finnemann, Peder Voetmann Christiansen.
Downward Causation: minds, bodies and matter.
Aarhus University Press. 2000