A framework for the analysis of evolutionary algorithms
Randall D. Beer, Leslie D. Picardo · 1996
Evolutionary algorithms are being used as a tool for function optimization. Several algorithms have been proposed differing in their genome representation, reproduction operators and selection procedures. These algorithms are nonlinear, discrete, stochastic dynamical systems and we need practical guidance to select the best algorithm for any given problem. This thesis describes a common framework for the analysis of evolutionary algorithms. We propose a useful metric to compare the performance of evolutionary algorithms on any given cost function. We introduce a sequence of simple evolutionary algorithms and derive analytical expressions that allow us to compute the performance of these algorithms on any cost function. We present results for the performance of this sequence of algorithms on two families of multimodal one-dimensional cost functions. We also present a useful software toolbox of C++ classes that allow the construction of a variety of evolutionary algorithms.