Fake it 'till you make it: How advancing AI Imitation Learning could contribute to the development of Artificial General Intelligence

Abstract

Research into Artificial Intelligence has sought to create a human-like machine, referred to as Artificial General Intelligence (AGI). Reinforcement Learning (RL), Artificial Neural Networks, the more recent Deep and Convolutional Neural Networks and, for this thesis most importantly, Imitation Learning (IL) have all been developed for this purpose. But until now, recent AGI-research, e.g. Google DeepMind, underlines Reinforcement Learning as the key to developing AGI, as this is how biological systems learn as well. This thesis, however, argues that this current paradigm unjustifiably neglects IL for its inferior specific task performance. This literature study illustrates the latest advancements in RL, then compares these systems with pure IL and mixed approaches. Next is a discussion on the essentiality of IL in our only functional example of AGI: the human mind. Finally, all chapters are brought together in a comparison between artificial and human IL, and it ends on a brief summary of how IL currently contributes to RL-systems in the field of AI.