Accelerating Evolutionary Algorithms Using Fitness Function Models

An optimization procedure using empirical models as an approximation of expensive functions is presented. The model is trained on the current set of evaluated solutions and can be used to search for promising solutions. These solutions are then evaluated on the expensive function. The resulting iterative procedure is analyzed and compared on test problems and shows fast and robust convergence. In particular, implementation issues like local modeling of the data, parallelization of the function evaluation and avoiding premature convergence are discussed.