How to Learn Quickly: An investigation of how to optimally train deep neural networks and its implications for human learning

This paper investigates the speed at which a deep neural network learns to play Tic- Tac- Toe as a function of the player's opponent's level of play. The goal of this work is not to find a faster deep learning algorithm or architecture, but rather to gain insight into the nature of adversarial learning and training. Do we learn faster by playing against an opponent of slightly worse, identical, better, or grandmaster ability?