Virtual Genetic Evolution

Sleeperservice writes "This story at New Scientist describes how, using cell simulation in computers, evolution can be simulated. How long until we can work out what the DNA sequence for a Dragon should be I wonder?"

Comments

GIGO

[Ben Jackson]

Simulated environments are just too complex.

And genetic algorithms, knowing nothing of your problem domain, tend to find solutions that capitalize on facets of your environment that you hadn't even considered. Two examples from my experience:

I had a population of simulated organisms competing in a shared 2d grid for food, which appeared in a pile at a random location when the old food was depleted. While the organism had basic looking/moving operations to rely on, invariably some would discover that with enough organisms, the food moves enough that you can survive by just looking around until the food is in your line of sight, and then jumping on it. My arbitrary decision to place the food randomly formed the basis for an *entire species* of organisms (which didn't fare too well when some got smarter).

These same organisms used a stack to do their thinking. Looking and eating produced values, which could be used for simple branching. Out of sheer laziness, I designed the stack to allow infinite pops off an empty stack which would return false, and infinite pushes on a full stack which would discard the values. One memorable run produced a dominant species which relied on this stack behavior to implement COUNTING! It intentionally (well, purposefully) left crud on the stack in a main loop, relying on the filled-stack behavior to detect a certain number of iterations. The stacksize and the arena size happened to be comparable, and this is how it determined when to turn.