Tuning Evolutionary Algorithms in High Dimensional Classification Problems

Evolutionary algorithms have been applied to high dimensional classification problems in order to look for the optimal set of predictive features. Crucial to the success is a proper configuration of these algorithms, in terms of fitness function, genetic operators and parameters settings since small changes in requirements can lead to completely different results. Tuning an evolutionary algorithm, as described in this paper, constitutes of choosing efficient crossover and mutation rates along with the calibration of the population size. For parameter tuning, the paper considers a very flexible evolutionary method where a Genetic Algorithm (GA) promotes the selection of better solutions and valuable results from different ranking methods provide a way of guiding the GA toward higher-accuracy. Extensive experiments on the classification of a public micro-array dataset allow highly effective tuning that also has the benefit of revealing important correlations between the GA parameters.