Comparative Study of the Starting Stage of Adaptive Differential Evolution on the Induction of Oblique Decision Trees

This study describes the application of four adaptive differential evolution algorithms to generate oblique decision trees. A population of decision trees encoded as real-valued vectors evolves through a global search strategy. Three schemes to create the initial population of the algorithms are applied to reduce the number of redundant nodes (whose test condition does not divide the set of instances). The results obtained in the experimental study aim to establish that the four algorithms have similar statistical behavior. However, using the dipole-based start strategy, the JSO method creates trees with better accuracy. Furthermore, the Success-History based Adaptive Differential Evolution with linear population reduction (LSHADE) algorithm stands out for inducing more compact trees than those created by the other variants in the three initializations evaluated.