Designing Evolutionary Algorithms

The essential idea of evolutionary problem solving is quite simple. A population of candidate solutions to the task at hand is evolved over successive iterations of random variation and selection. Random variation provides the mechanism for discovering new solutions. Selection determines which solutions to maintain as a basis for further exploration. Metaphorically, the search is conducted on a landscape of hills and valleys (see figure 7.1), which is also called a “response surface” in that it indicates the response of the evaluation function to every possible trial solution. The goal of the exploration is most often to locate a solution, or set of solutions, that possesses sufficient quality as measured by the evaluation function. It’s not enough to simply find these solutions, however, we need to find them quickly. After all, enumeration will find such solutions too, but for real problems we’d grow old waiting for the answers. The speed with which suitable solutions can be discovered is in part dependent on the choices we make in determining the representation of trial solutions, the evaluation function, the specific variation and selection operators, and the size and initialization of the population, among other facets. The key to designing successful evolutionary algorithms lies in making the appropriate choices for these concerns in light of the problem at hand.