Most Evolutionary Algorithms are Simple

My newest idea for implementing a library for GPM, GA, GP, GEP, particle swarm type systems is that there is just no way to cover all cases reasonably. We can't really say anything in general about what we are evolving, how the evolution occurs (which operators, structures, etc), or anything really. We can't (without losing generality) fix the structure of the algorithm, or what problem it solves, or what the individuals look like, or what the population looks like.

Therefore I am relying on the relative simplicity of Evolutionary Algorithms and rather then providing a reusable library that forces the user to only express certain types of EAs, I'm just providing code for the ones I want to do, and even for those ones I'm just constructing the algorithms and data structures pretty much directly, with very little abstraction. The idea is that the user just creates everything from scratch every time with help from the library, but not simply by composing structures in the library, as that doesn't seem feasible.

The best way I can think to help users (myself) create EA systems is to make sure the library stays out of the way. EAs can come in any shape or form, and the important thing is to fill out design spaces that might be useful in as lightweight a way as possible. Don't force the user to redo *everything* just so that they can vary some part of the system that usually doesn't vary, and make it easy to construct a new method using only the relevant parts of my library.

This is kind of a concession to the complexity of the problem- I have to start running tests instead of trying to get the program to be as pretty as possible. If this means poor code, I will do my best and hope to fix it when I come up with something better. I'm not unwilling to start over if there is some nice way of doing this.