I'm going to use the phrase as a jumping off point to talk about GAs. Genetic Algorithms are explicitly designed to imitate the successful process of optimization seen all around us.
Simple GAs try to tear down the bells and wistles of real biology and use just the core ideas of evolution. Take a finite population, evaluate the fitness of each member. Preferentially select higher fitness members to reproduce or be copied forward. Allow some source of variation to create new population members during reproduction. The most common reproductive operators are mutation and crossover.
A choice of population size, selection algorithm, operators and representation that all work together will take even a simple GA a long way. But GAs are sometimes criticised as being unable (even in principle) to replicate the diversity we see in the natural world.
Most GAs working on an optimization problem are like a microbe evolving drug resistance in a test tube environment. No one is even attempting to recreate the variety of species and niches of real life.
Here are some things that are sometimes brought up as examples of how GAs can be made more "biological", in the belief that this will somehow make them better:
- haploid genomes
- inversion operator
Now it may be that there are specific optimization problems where each of these tricks might be helpful, but none of them has been shown to be extremely useful.
Notice that these ideas focus on the nitty-gritty of DNA. I have my own list of things I think might be important to bring in to our GA models from biology, but these aren't on my list. Instead, here is my list:
- embodiment - in time and space
- coevolution - a fitness measure in which the other members of the population are explicitly considered, including the possibility of parasitism
- development - genetic cascades, and some form of lifetime fitness measure
- self regulation - in the form of hormones or neurons
- sexual selection
I'm much more interested in the possibilities of a boost from ecology than biology. It may be that one day, very powerful GAs will drop some things that are now considered "essential" such as the population itself! Estimation of distribution algorithms are a step in that direction. Until then, I think we do have a lot to learn from Mother Nature, but I think we need to look in some different directions.