Dynamic environments can speed up evolution with genetic programming

We present a study of dynamic environments with genetic programming to ascertain if a dynamic environment can speed up evolution when compared to an equivalent static environment. We present an analysis of the types of dynamic variation which can occur with a variable-length representation such as adopted in genetic programming identifying modular varying, structural varying and incremental varying goals. An empirical investigation comparing these three types of varying goals on dynamic symbolic regression benchmarks reveals an advantage for goals which vary in terms of increasing structural complexity. This provides evidence to support the added difficulty variable length representations incur due to their requirement to search structural and parametric space concurrently, and how directing search through varying structural goals with increasing complexity can speed up search with genetic programming.