These claims about where evolutionary methods are useful need to be qualified.<p>1. They need to have a domain where a given candidate solution can be quantitatively measured, or alternatively where you can induce a relative ordering. This could possibly require human input ("rank these three pictures from best to worst"), but you need to at least have the ability to say that one solution is more fit than another.<p>2. A candidate solution's fitness needs to be at least somewhat continuous. Ideally there needs to be a gradient of success and opportunities for incremental quality improvement with corresponding incremental fitness improvements. Evolutionary methods cannot solve needle-in-a-haystack problems.<p>3. They're of greatest utility where human ingenuity is of the littlest value: i.e., where the problem domain is little understood. For example, finance, and, apparently, things like radar design: black arts where no one really knows what the fuck. Ths should help mullify claims about human verification and the human element of design: these things work best where a lifetime of experience doesn't buy you much and you kind of have no other choice.<p>4. As with any search heuristic, you face an eternal battle between exploration of the search space and exploitation of it: you can exhaustively explore one part of the space, but that will generally come at the cost of ignoring alternative solutions. This manifests itself as a struggle against premature convergence. Depending on how much domain knowledge (an often-overlooked component) you can incorporate in, this may not be a problem (you can modify your fitness criteria to automatically avoid settling upon certain undesirable parts of the search space).<p>5. Finally, always keep in mind that this is not "AI". This is not trying to model human cognition. The stuff about being inspired by natural selection is true at first, but when you get more sophisticated you'll deviate more and more from this. Hybrid solutions are common; for instance, you may have a normal stochastic evolution scheme at first, but then at the end greatly slow down the rate of learning and at the end when you've settled upon a satisfactory solution you may wish to incorporate traditional local search methods, which really have no biological inspiration, it's just that they just freakin' work, so you use it because you care more about <i>solving your problem</i> than constantly aping nature.<p>Biological or natural inspiration can be useful, but only in the context of practical problem solving, which is what evolutionary methods have always been about. The first widely read book on genetic algorithms was written by an oilfield roughneck with a PhD in civil engineering and the field is still kind of a pariah in academia because of its relative lack of formal foundations. All this is disregarded in the name of practicality, which is an attitude, flawed or not, much closer to what prevails in the real world.