Direct evolution of hierarchical solutions with self-emergent substructures

Linear genotype representation and modularity have continuously received extensive attention from the genetic programming (GP) community. The advantages of a linear genotype include a convenient and efficient implementation scheme. However, most existing techniques using a linear genotype follow the imperative programming language paradigm and a direct hierarchical composition for the functionality of the solution is underachieved. Our work is based on prefix gene expression programming (P-GEP), a new GP method featured by a prefix notation based linear genotype representation. Since P-GEP uses a functional language paradigm, its framework results in natural self-emergence of substructures as functional components during the evolution. We propose to preserve and utilize potentially useful emergent substructures via a dynamic substructure library, empowering the algorithm to focus the search on a higher level of the solution structure. Preliminary experiments on the benchmark regression problems have shown the effectiveness of this approach.