Model-Based Reinforcement Learning in Atari 2600 Games
Abstract
Humans are able to develop an understanding of the dynamics of a video game through experience playing the game. They can utilize this information to predict future screens of gameplay and thus become better players. Researchers have created software agents that do not attempt to predict future screens of gameplay, but are able to achieve success in Atari 2600 games through reinforcement learning. Researchers have also created methods for future screen prediction within Atari 2600 games, but have not demonstrated these methods could allow a software agent to achieve gameplay success. This document will focus on the intersection of reinforcement learning and future screen prediction within Atari 2600 games. This work will highlight the challenges related to this intersection that are not often presented in the current AI literature.
Description
Franklin and Marshall College Archives, Undergraduate Honors Thesis 2017
Collections
- F&M Theses Collection [322]