|
Is there any modification of GAs, which would be just purely based on mutation as an evolution strategy? Something similar to bacteria evolution? |
|
This idea is described (and well studied and better theoretically understood than genetic algorithms with crossover) by the "replicator dynamic". Basically, you imagine you are optimizing within a space with a fitness function based on that space. You create a bunch of points scattered through this space, kill the ones that have a bad fitness, and replicate the ones with good fitness functions in proportion to how good their functions are (maybe with noise or some other criterion to create variation among the replicated ones). This works well in general (the particle swarm algorithm is described by this), and guarantees convergence on convex fitness landscapes. |
|
I'm not sure your question is too clear. For starters have a read of this, if that doesn't help, you might need to expand your question. 1
I think the question is about using evolutionary algorithms that don't use sexual reproduction/gene mixing.
(Jan 03 '12 at 17:32)
Rob Renaud
...so just mutate the individuals in the population with a high fitness function, but don't use other genetic operators like crossover? If so, then the answer to the OP would be sure.
(Jan 03 '12 at 18:30)
amair
|
|
As amair notes, there are several optimization techniques that rely (mostly) on mutation. Simulated annealing and particle swarms are arguable examples, although SA usually considers just one individual at a time, and with particle swarms there is some feedback from the overall best individuals back into the general population. As an aside, though, you're incorrect if you're assuming that, because bacteria reproduce asexually, their genomes only change through mutation. There is plenty of lateral gene-swapping that goes on. I'm probably going out on a limb, here, but this suggests to me that mutation-only systems are possibly inferior, since even in the absence of sexual reproduction, evolution has found a way for microbes to mix and match. |
|
According to the theory of punctuated equilibrium (Gould and Eldridge), cross-over helps average out the population. Thus, in a simple model, mutation would be the main operator. I've also see claims that cross-over for GP doesn't help much and that just mutation is enough. Of course, in real life you have a bunch of other factors. Just because a gene is present doesn't mean it gets activated (penetrance). And just because it has penetrance, the degree to which it gets expressed can vary. Both factors can be affected by the environment. Not to mention how genes can interact with each other, etc. |
