Applying image segmentation to deep reinforcement learning in video games
LE3 .A278 2022
2022
Lee, Greg
Acadia University
Master of Science
Masters
Computer Science
Deep Reinforcement Learning (deep-RL) has been used to create agents that can outperform top human players in game domains from chess to modern video games. The models using the most general techniques in this field apply image preprocessing techniques, such as downsampling, that remove key information and non-key information in equal measure. This thesis proposes a novel method of generating image segmentation datasets in domains which can be simulated, including video games. The intention behind this is to enable the creation of image segmentation models that can preprocess image state input to a deep-RL model in a more intelligent manner by removing non-key information from the images. In testing these methods, it was found that image segmentation improved the sample efficiency and reduced the learning time of a deep-RL agent in Pong. Furthermore, it was found that a segmentation model using the proposed automatic method trained on Super Mario Bros., an environment where deep-RL is still behind human ability, outperformed a model using more traditional manual labelling methods in segmentation accuracy and dataset creation time.
The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.
https://scholar.acadiau.ca/islandora/object/theses:3872