Evolutionary Dynamics Effects Account for the Improvement Caused by R-Augmentation

Previous work has found a method for augmenting a genetic algorithm (GA), referred to as R-augmentation, that produces better-scoring results earlier in evolutionary time. R-augmentation works by using an information-theoretic measure quantifying mental representations (denoted as R) as an additional contribution to the fitness function of the GA. It has been shown that this method improves the performance of the GA by encouraging the evolution of high-performing artificial agents controlled by Markov Brains or Recurrent Neural Networks. Different mechanisms could explain this phenomenon. Here we demonstrate that the majority of the improvement is caused by R-augmentation altering the evolutionary dynamics of the GA, reflected in the changed movement of mutational operators through the fitness landscape.