On genetic algorithms and discrete performance measures

A relation exists between the manner in which a statistical model is developed and the measure employed for gauging its performance. Often the model is developed by optimizing some continuous measure of performance, while its final performance is assessed in terms of some discrete measure. The question then arises as to whether a model based on the direct optimization of the discrete measure may be superior to or significantly different from the model based on the optimization of continuous measure. Some Artificial Intelligence parameter estimation techniques allow the optimization of discrete measures. Genetic Algorithms constitute one such technique, and therefore, allow for an examination of this question. Here, one type of genetic algorithm is employed to optimize three discrete performance measures of a parametric model for the prediction of hail. A more conventional technique is then employed to optimize the same discrete measures. The former outperforms the latter. In other words, the direct optimization of three discrete measures via genetic algorithms yields better fits to the data than alternatives requiring the intermediate step of optimizing a continuous measure.