I want to make sure we're all on the same page, before we start digging through source code. The way simulated evolution works is you make a population—maybe totally random, maybe not—and call it generation 0. You then run some fitness function on the members of the population. Fitness may be defined in terms of some known absolute or as a relative value between members of the population. The fittest members of the population are then taken to form a subset of parents. Many algorithms take two members, but the sky's the limit and your species could well require six individuals for reproduction. That's what we're after—reproduction.

The individuals of the population are non-destructively recombined to form children. The population is then mutated, changing genes in an individual's genomes with some domain-specific probability. This may be all members of the population or just children, or just parents, or whatever you'd like. There's no perfect answer here. After mutation, the parents, their children, and any non-reproducing members of the population that did not die during fitness evaluation are then promoted to generation 2. The cycle continues:

• Fitness evaluation
• Parents reproduce
• Population is mutated

Exactly when the cycle stops is up to the programmer: perhaps when a fixed number of generations have passed, perhaps when an individual that meets some minimum threshold of fitness has evolved.

Simulated evolution is not magic. Much like QuickCheck, which uses randomness to probe programs for property violations, this algorithm is probing a problem space for the most fit solution. It's an optimization strategy, mimicking the biological process. As you can probably infer from the description I just gave, there are a lot of knobs and alternatives to the basic approach. We'll be putting together a straightforward approach here but the you are warmly encouraged to make your own modifications.