Digital organisms: Tierra, Avida and Polyworld

Long before this project, several systems put self-replicating entities into a computer and let evolution run without supervision. Their results, and their limits, are the background against which any new attempt should be read.

Tierra

Tom Ray's Tierra, from 1991, gave programs a virtual computer of their own and let them compete for processor time and memory. Ray wrote one self-replicating ancestor by hand and then stopped intervening. What followed was not planned by anyone: shorter replicators appeared that copied faster, then parasites that had dropped their own copying code and used a neighbour's, then hosts resistant to parasites, then parasites that got around the resistance.

Nothing in Tierra rewarded parasitism. It appeared because it paid, in a world whose only rules were about instructions, memory and time. It remains one of the cleanest demonstrations that an ecological relationship can arise without being specified.

Avida

Avida, developed from the mid 1990s at Michigan State and still in active research use, put digital organisms on a grid and rewarded them with processor time for performing logic operations. That design made it a genuine experimental instrument rather than a demonstration, because populations could be replicated, treatments compared, and every ancestor of every organism recovered exactly.

Its best known result, published in Nature in 2003, traced the evolution of a complex logic function and showed it was built from simpler functions that had been useful for other reasons first. When the intermediate steps were denied any reward, the complex function never appeared. That is a direct experimental result about how complex features arise, of a kind that is extremely hard to obtain from the fossil record.

Polyworld and the embodied line

Larry Yaeger's Polyworld, from 1994, took the opposite trade. It put agents with neural controllers and evolved body properties into a two-dimensional world with food, vision and physical interaction. Rather than measuring a logic function, it asked what happens to the complexity of a population's neural organisation when the world is a place rather than a program.

This embodied line, which also includes work such as Karl Sims' evolved virtual creatures and a long tail of ecosystem simulations, gives up the clean experimental control that Avida has. The results are harder to reduce to a single measured quantity, and the systems are far more expensive to run.

Where Emergent Life sits between them

Emergent Life is in the embodied line, and it inherits that line's difficulty: a rich world is harder to draw a defensible conclusion from than an abstract one. The project's response is to borrow the discipline from the other side. Runs are exactly reproducible from a seed, rulesets are frozen for the life of a universe, and results are pre-registered before the run that tests them. How that works is on methodology and determinism and reproducibility.

The other difference is the absence of a reward. Avida pays organisms for computing logic functions, which is what makes its experiments so sharp. Emergent Life pays for nothing. Organisms spend energy, find more or die, and reproduce as an action of their own, which is described in energy, metabolism and death. That removes the sharpest available experimental lever, on purpose, and makes progress correspondingly slower.