What is artificial life?

Artificial life, usually shortened to ALife, is the study of living processes by building them rather than by dissecting them. Where biology examines the life that exists, artificial life asks what any life would have to do, and then tries to make something do it.

The short definition

The field was named at a workshop in Los Alamos in 1987, organised by Christopher Langton. His framing has held up well: biology studies "life as it is", and artificial life studies "life as it could be". The point is not to imitate a particular organism. It is to find which properties of living things are consequences of chemistry and history, and which are consequences of being an organised, self-maintaining, reproducing system at all.

That gives the field an unusual test of success. A simulated bird that flies convincingly is a graphics achievement. A simulated population that was never told about flight, and in which something flight-like appears anyway because it pays, is an artificial life result. The difference is whether the interesting behaviour was put in or came out.

The questions the field is actually asking

A few recurring problems define most of the work:

• Origins. How does a system that copies itself arise from one that does not?
• Open-endedness. Why does biological evolution keep producing genuinely new kinds of thing, while most simulations settle down and stop? See open-ended evolution.
• Levels of organisation. How do individuals become groups that behave as individuals: cells into bodies, bodies into colonies?
• Autonomy. What separates a system maintaining itself from a system merely being maintained by its surroundings?

Most published ALife work makes progress on one of these in a deliberately narrow setting. The broad, everything-at-once world is rarer, because it is much harder to draw a defensible conclusion from one.

What artificial life is not

It is not the same as machine learning, though the two overlap and borrow from each other freely. A machine learning system is normally given an objective and optimised against it. An artificial life system is normally given an environment and a survival constraint, and whatever the population does about that is the result rather than the target. When ALife uses learning methods, it usually uses them as the mechanism inside an organism, not as the goal of the experiment.

It is also not a claim that a simulation is alive. Whether a sufficiently detailed simulation of a living process would itself be living is a genuinely open question in the field, argued in both directions, and it is not settled by any result on this site.

Where Emergent Life sits

Emergent Life is a soft ALife world with no fitness function. Nothing in it scores an organism or ranks a population. Organisms sense generic quantities, act through generic channels, spend energy, and either reproduce before they run out or they do not, and selection is whatever that leaves behind. The rules are frozen for the life of a universe, so a result cannot be produced by quietly retuning the world until it appears.

What that buys, and what it costs, is set out on methodology: no objective means nothing can be optimised toward a demonstration, but it also means progress is slow and many runs produce nothing worth reporting. Those runs are written up too, on research notes.