VizDoom AI via Deep Reinforcement Learning

Aaron James


Supervised by Frank C Langbein; Moderated by Steven Schockaert

In this paper, I will adapt an existing algorithm to the ViZDoom visual deep reinforcement learning library and try to find optimal algorithmic settings for some default scenarios provided by the library under the PPO and A2C algorithms. I will attempt to further this into playing The Ultimate DOOM once these settings are found. I will follow existing novel methodologies such as transfer learning, curriculum learning, and reward shaping to allow the machine learning algorithm to develop an understanding of its environment more quickly and to enter complicated scenarios with existing knowledge from previous, more basic scenarios.

Initial Plan (07/02/2022) [Zip Archive]

Final Report (13/05/2022) [Zip Archive]

Publication Form