By James Mapes

Temporal Difference Learning, Neural Networks and Rook

April 2002

Advisor: Dr. Jeff McKinstry

Abstract:

Objective: To apply a temporal difference learning algorithm to a simplified version of the card game Rook. The goal was to automatically develop a strong Rook agent with little or no knowledge programmed into the system; instead the agent would learn from playing against opponents that played randomly.

Experiments: A neural network was trained to be an evaluation heuristic for a simplified version of Rook. The network used a temporal difference learning algorithm to train during self-play. The results of the error from the evaluation heuristic and the average scores were recorded.

Results: The average scores of the game playing agent using the neural network improved significantly with training. The error of the evaluation heuristic also decreased.

Conclusion: The results of this research indicate that temporal difference learning with neural networks can be used to create an evaluation heuristic for complex card games.

 

By Catherine Pfeiffer

How Do Graphs Sound?

April 2002

Advisor: Dr. Maria Zack

Abstract:

This project looks at one of the connections between mathematics and music. Specifically, it looks at how various graphs sound and how various translations, such as rotations, reflections, and slides, change the sound of the graphs. Random walks on graphs are also "played". In addition, this project compares atonal music to tonal music by taking various atonal and tonal songs, graphing them, and then comparing the different shapes of the graphs.