RBH

KeithHarwood wroteThat's not always the case. My company's EAs, for example, do not have a specified goal. We have multi-variate relative fitness metrics, but specify no goal as such for the evolving entities beyond 'survive and compete successfully in the environment given the several components of the fitness metric.' In fact we have not the slightest idea what the goal state might look like, and evolve populations composed of species with different properties rather than single solutions. Further, we don't control the (complex) selective environment -- that's piped into the EA shell from the outside world, and we have no idea what the fitness landscape looks like. We're not using the EA to optimize a solution but rather to generate populations of critters artificially evolved against the piped-in environment which then operate to produce (hopefully!) adaptive behaviors transmitted to the outside world to drive real-world processes. Our EAs never terminate, because the external environment piped in to them is itself dynamic.

RBH