Using Reinforcement Learning to Train Generative Adversarial Networks for Image Generation

Image generation using machine learning techniques is not a new concept and the most popular class of models is the Generative Adversarial Networks, known as GANs. This paper will investigate a novel method of using deep reinforcement learning in conjunction with the GANs framework to train neural networks to produce images. Two reinforcement learning algorithms - the Advantage Actor-Critic and the Twin Delayed Deep Deterministic Policy Gradient algorithms, are applied in this preliminary work. Several models were experimentally trained using varied hyperparameters and the quality of the resulting generated images shows promise and presents a baseline for further study on this problem.