Rethinking multilevel selection in genetic programming

This paper aims to improve the capability of genetic pro-gramming to tackle the evolution of cooperation: evolving multiple partial solutions that collaboratively solve struc-turally and functionally complex problems. A multilevel genetic programming approach is presented based on a new computational multilevel selection framework [19]. This ap-proach considers biological group selection theory to encour-age cooperation, and a new cooperation operator to build solutions hierarchically. It extends evolution from individu-als to multiple group levels, leading to good performance on both individuals and groups. The applicability of this ap-proach is evaluated on 7 multi-class classification problems with different features, such as non-linearity, skewed data distribution and large feature space. The results, when com-pared to other cooperative evolutionary algorithms in the lit-erature, demonstrate that this approach improves solution accuracy and consistency, and simplifies solution complex-ity. In addition, the problem is decomposed as a result of evolution without human interference