Evolution of evolvability for neuroevolution

Scientists have been in awe of the powers of evolution since comprehending the fundamental principles of natural selection. As Darwin eloquently put it, nature has given rise to “endless forms most beautiful”. However, beyond the mesmerizing beauty of nature, lies an even more fascinating aspect - the creation of human intelligence. Although AI research has produced impressive outcomes in the past decade, the techniques employed to accomplish these outcomes diverge considerably from the natural process of brain evolution. The primary dissimilarity stems from the fact that nature employs the evolution of DNA code, which encodes the instructions for brain development, whereas mainstream AI represents all brain parameters directly. This disparity is noteworthy as the indirect representation utilized by nature confers upon it a potent ability to enhance evolvability over time, a feature that is absent in direct representation. This potential stems from the ability to choose instructions in a manner that renders the brain resistant to change in certain directions while facilitating easy modification in other directions. The present thesis delves into the potential of leveraging indirect encoding and evolvability during the neural network evolution process. To this end, we propose two algorithms, namely Quality Evolvability ES and Evolvability Map Elites, which enable direct selection for evolvability. We conduct an evaluation of the efficacy of these algorithms in the context of robotics locomotion tasks. Additionally, we investigate the necessary conditions for indirect encoding to be useful for learning and conduct experiments to verify our hypothesis in the domain of image recognition tasks