Symbolic Regression with Sampling

In this paper a way of improving the performance of genetic programming (GP) for regression tasks is presented. In general, most of the execution time is consumed during the evaluation step of an individual. Hence reducing the number of samples which are evaluated during the learning phase of the algorithm significantly reduces its execution time. A reduction of the available training samples might hamper the algorithm in its capability to learn the desired correlation. For this reason our approach evaluates each solution only on a randomly chosen part of all training samples, which is selected before the evaluation step. In the result section runs with different parameter settings of our approach and traditional genetic programming algorithms are compared regarding the solution quality and execution time to each other.