Artificial Life — Some Thoughts
Thinking about artificial life, written in 1996.
Artificial Life, some thoughts about
Artificial Life, some thoughts about
See also Santa Fe Institute for ALIFE research
Further explorations:
A look at emergence of atoms in our universe
"The Outer Reaches of Life" a book by John Postgate
Cambridge University Press 1994, ISBN 0-521-44010-6
"Biology" by Scott Foresman 1985, ISBN: 0-673-13735-X
"Organic and Biological Chemistry" by John R. Holum
1986 John Wiley & Sons Inc ISBN 0-471-83114-X
"Fluid Concepts and Creative Analogies" a stupid book by Douglas Hofstader
1995 Basic Books / Harper Collins Publishers Inc. ISBN 0-465-05154-5
"Atoms of Silence", Hubert Reeves
"Introduction to Artificial Intelligence", Philip C. Jackson, Jr.
"Life in the Universe" Oct 1994, Scientific American.
"Genetic Algorithms", June 1994, Computer Magazine, IEEE Computer Society.
Our universe seems to celebrate a perverse tendency for the simplest of things to become inextricably tangled up in their determined attempts to seek the lowest energy state. It seems to promote self-organizing behaviour in the face of (or perhaps because of) entropy.
What kind of rules does a universe need to promote spontaneous self-organization of its constituent elements? Or to phrase it another way: The universe tends towards complexity, and may tend towards life itself. If so, why is life part of the fitness criteria of the universe? What "profit" is there in favouring life?
An article in Scientific American, February 1995 suggests that self-replicating RNA might have been the original basis for life. Self-replication seems to be an underlying mechanism, an "engine" permitting other processes such as evolution, and this grand march towards complexity that we see in the universe around us.
Life forms seem to very carefully preserve and exploit self-organization and self-replication. Perhaps by studying life and the way it exploits the rule space of its universe it may be possible to illuminate both the "how" and "why" of the universe.
I propose that we simulate a primordial soup that can exhibit many layers of emergent behaviour autonomously, including the spontaneous autocatalysis of life. This would be a nice contrast to Tierra which presupposes aliveness, and would help us understand what laws are at work in our world.
Why bother creating a new simulation when other simulations already exist?
In fact there doesn't seem to be justification to create any simulations. Given that all computers can emulate all other computers then a sufficiently large version of Conway's Life should provide the same gratification as a custom simulation with custom rules.
But then why not do all computation with variations of Conways Life? Why have spreadsheets or UNIX kernals? Obviously modern computer architectures favour certain kinds of instructions and operations that Conways Life fails to exploit. Tailoring a custom application to better use those resources will give us results in a much shorter period of time.
Interestingly enough, the universe of Conways Life can be viewed as one in which the rules were insufficiently rich for the limited information and computational space. The chances of Conway critters exhibiting self-organizing behaviour and emergence in 8 megabytes of memory is vanishingly small. The environment itself is a barren, sterile ground able to support only the simplest effects.
Although admittedly, there are some very interesting sparse-matrix hashed versions of Life, which compress duplicate blocks of life-space. Vastly larger simulations are possible than one might think.
Can the rules of an artificial simulation be truly independant of our outer universe?
Will our attempts to create an evolving artificial environment succeed because the our universe is so very conducive to spontaneous self-organization. The implication being that somehow there might be properties of our universe that pollute any experiment we try to create. Classical information theory ought to be able to answer this one. The answer seems to be that a simulation can be independant, except that it won't be able to simulate areas of the real universe wherein there exist singularities or parts of the system unknowable to entities within the system.
If a virtual simulation is itself a true computer then it should be able to emulate any other computer and thus presumably be able to exceed the limitations its parent may have at the raw instruction level. If the parent computer lacks the ability to manipulate bits then virtual computer can use virtual bits. Of course the virtual computer will never be able to access physical memory outside the boundaries of the parent - but in a truly virtual sense the limits themselves will be virtualized.
Another problem is the definition of life itself; what kind of definition should we have?
For the purposes of this study let us define life as something which is able to self-replicate and is based on emergent behaviour; having interactions based on rules derived from an assembly of simpler parts. Although the two criteria are overly strict - any creatures found in that domain would probably meet most criteria of life - and a "life-watcher" program can use this criteria to report to us the first moment of life in our simulation.
How does our universe work? Can we re-create the behaviour we are seeking?
There is the implication that our real universe is an evolved entity. If so, how does our universe express its genome? How does it give birth to children and pass on its traits? If the universe is cyclical then the further implication is that time is outside the universe. If not, and time is just one more attribute, then are all other universes being manifested in parallel with the current one? Which implies that they might interact at the rule level; having fickle rule boundaries. This may be fundamentally different from from what we are doing and may relate to the propensity towards complexity in a way our simulation won't mimic.
Can we even safely assume there is some underlying rule-set in our universe? That the rules themselves aren't buffeted by the particles nearby? It would seem that there has to be one turtle at the very bottom but all our cultural prejudices will be strongly evident in anything we create. We may find the gold which we set out to seek, missing the diamonds along the way.
We don't even know if a serialized cellular autonoma can be truly the same as a parallel simulation. If our universe is indeed parallel and a serial computer cannot mimic some parallel processes then our simulation may not serve as a good model for understanding real complexity and life.
It seems like the best simulation is one that is as simple as possible; able to seek its own ideal rule space from a very simple substrate; moderately tailored towards our goal of spontanteous self-replication.
For a universe to evolve it needs to be able to quantify itself. How can we quantify the complexity of our universe? Here are two ideas:
First I propose that we consider the idea of differentiating between "complexity" and "noise" by bringing the observer into the picture. A "complex" object is both an interpretor-engine and mass-of-data paired together. Noise is that which has no interpretor-engine.
Secondly, the rules of a finite universe and how they act on elements in the universe may help or constrain the propensity for self-replication, and the degree to which self-replicating life forms can evolve. If the rules of a universe are very simple, and thus the kinds of interactions that elements in that universe can have, then more elements are needed to preserve the same quantity of information. Rules and interactions multiplied by elements yields a total information volume?
At an emergent phenomenological level there seem to be new layers of rules that are in a sense selected for:
An emergent phenomena, such as say an amino-acid depends upon the presence of certain constituent elements. In a sense, the elements become rules in themselves which mediate the amino-acids effects. The amino-acids are found in a subset of the possible rule space where the rules are conducive to its effects. For example, by relying on carbon the amino acid is able to maximize it's bonding ability. We should consider rules to be not only at the fundamental level of the universe, but as artifacts of emergent phenomena, existing at many levels.
Although presumably the lower level rules would still be operating so a total quantifiable information volume should be computable; although it may be a constant.
This may indeed allow a method by which the kinds of rule-spaces that govern universes might be selected for. There may be a way to evolve universes which favour emergent phenomena; perhaps not at the most fundamental rules level, but on the first derived level from that: Ideally each iteration of our simulation will be able to modify its parameters to be ultimately better suited for emergent phenomena.
Obviously we need more information, so let us start by examining our own universe.