Hybridizing Lévy Flights and Cartesian Genetic Programming for Learning Swarm-Based Optimization

Cartesian Genetic Programming is a well-established version of Genetic Programming and has meanwhile been applied to many use cases. The case of learning swarm behavior for optimization recently showed some fitness landscape characteristics that make program evolution harder due to the intrinsic barrier structure that is hard to pass by using standard mutation. In this paper, we explore possible improvements by replacing the standard uniform mutation by Lévy flights when training with a \((\mu +\lambda )\) -evolution strategy. We demonstrate the superiority of the new variation operation for training instances of the optimization learning problem and compare success rates and minimal computational effort.