Genetic algorithms are a method of self adaption by means of evolution.
Basically, take a group of characteristics which define your problem solver. The problem solver could be an ALife agent, obviously with a variety of control methods. For this example I'll use a neural network control method.
The agent has a number of inputs and outputs, inputs including information about the surrounding area, such as number and proximity of predators, amount of food. The outputs are the agents effectors, which include moving in a given direction, the speed of motion, eating or defending itself.
With a traditional neural net architecture you need a training set. A group of inputs with their correct outputs, to allow the neural network to setup the weighting of the connections between the neurons.
With genetic algorithms you would create a hundred agents with random networks of randomly connected neurons. They would then be set free in the world and their actions monitored. A fitness function would be used to find the best ten agents. For this example you could simply use the last ten agents to die of starvation or from being eaten or the ones who managed to get the most food or the ones who evaded the predators best.
The point being that you have to decide what problem you're solving before you can decide how fit an agent is to solve the problem.
You would then take the ten neural networks of the best agents and breed them together, pairing them off so that each agent breeds with every other agent, creating 81 new agents (you could make more or less, your choice). By breeding I mean you would take the characteristics of the neural nets and meld them together using crossover and mutation. If the architecture of the neural net was fixed i.e. x inputs y outputs z hidden layers, then you would be simply choosing the weights from either the father or the mother or possibly averaging them out.
You could then randomly adjust some of the weights by increasing or decreasing them.
If you mutate them too much then you destroy the breeding side of the algorithm.
Ones the offsrping have been generated then you start the experiment over again with the new agents, finding your next ten best and continuing for hundreds of generations.
Hopefully you would eventually come out with the perfect agent for your testbed.
No guarentees though.
Does this answer your question?