Student research opportunities
Building Agents that Learn to Play Games
Project Code: CECS_45
This project is available at the following levels:
Honours, Summer Scholar, Masters, PhD
Keywords:
Game Playing; Machine Learning; Reinforcement Learning
Supervisor:
Dr Scott SannerOutline:
Want to play a game?
Reinforcement learning (RL) is the task of learning to make optimal sequential decisions based on previous experience. Based on RL techniques, the world's best backgammon player TD-Gammon taught itself to play Backgammon via reinforcement learning techniques. One of the greatest obstacles to practical RL is finding the right features to use.
Goals of this project
In this project, you can examine your choice of automated feature induction methods for RL ranging from functional gradient boosting to artificial neural networks to kernel machines. Application domains can range from board games (Othello, Backgammon, Go) to online real-time strategy games like Stratagus or Wargus. Your goal is to design a learning algorithm that not only learns to perform its assigned task but also learns important structured features that help it play.
Requirements/Prerequisites
Basic knowledge of probability. Good Java programming skills a plus.
Student Gain
Understanding of the fields of Machine Learning and Optimal Sequential Decision-making.
Background Literature
"Reinforcement Learning", Rich Sutton and Andy Barto (1998).
Links
Scott Sanner's web pageBackground reading available online
