Augmenting a neural agent with an Oracle

Abstract

In this thesis, I approached deep reinforcement learning agents with the novel idea of augmenting the agent with another neural network called an "Oracle" to gain more insights about the game environment. The Oracle is trained through supervised learning and can be used in various ways with the agent such as a reward shaper or in a pipeline with the agent through which it can transform the original input state into a more enhanced input state with more information. Overall results were not positive as creating a good Oracle can be hard. Creating a better Oracle could possibly display promising results.