28th Annual Student Research Conference
A New Reinforcement Learning Algorithm for a Game-Playing Agent
Thanh D. Le
Dr. John Seiffertt, Faculty Mentor
Reinforcement learning algorithms are powerful tools to help computers understand and learn completely unfamiliar environments, but many of the algorithms in common use have weaknesses. On key weakness is inability to learn when the number of states used to represent the environment is large. Our research addresses this weakness by combining a reinforcement learning approach called Q-learning with another type of machine learning algorithm called Adaptive Resonance Theory (ART) to understand more about artificial intelligences and to create an agent that can learn and adapt to different environments By creating a new algorithm fusing Q-learning and ART we are able to demonstrate a more intelligent agent than the traditional methods allow. Our results are illustrated using the games Pac-Man and Bomberman where we show our algorithm results in a higher rate of success than does the standard approach.
Keywords: Reinforcement learning, Artificial Intelligence, Adaptive Resonance Theorem, Q-Learning
Topic(s):Computer Science
Presentation Type: Poster
Session: 5-1
Location: GEO-SUB
Time: 3:30