“Training Artificial Intelligence Agents to Play a Family of Combinatorial Games” by Maggie Haeussler, Lina Mo, and Sidney Wright (Advisor: Dr. Darin Stephenson)

This talk explores the application of various artificial intelligence techniques in developing strategy for combinatorial games. In this project, we study a family of deterministic 2-player games played on m by n grids with potentially some cells removed. These games involve players taking turns placing pieces on the board until the board is filled, after which sequences of pieces are scored based on length. Our research includes training AI agents with a variety of methods such as using tabular reinforcement learning, evaluating N-tuples of grid squares, and developing genetic training methods with artificial neural networks. We train these agents against a variety of non-learning agents and evaluate their performance against both non-learning agents and one another to assess the quality of decision-making. Additionally, we delve into the theoretical foundations of game theory and strategy specific to this family of games and explore how these strategies can be implemented in our AI agents.

“On the parameter tuning challenge for spectral clustering methods” by Valen Feldmann, Eli Edwards-Parker, and Irene Seo (Advisor: Dr. Gabriel Chen)

Spectral clustering is an exciting, attractive modern clustering approach, with many successful applications such as document clustering and image segmentation. It is not without challenges, however, as it has high computational complexity and requires parameter tuning. Ever since its introduction, much effort has been spent on making spectral clustering scalable (in both memory and speed) to large data sets whereas there is little work on parameter tuning in the context of spectral clustering (including the landmark-based methods). In this talk, we address the parameter tuning challenge of spectral clustering. Specifically, we propose a new criterion in the embedding space (where k-means is applied) for tuning two kinds of parameters: (1) the scale parameter used in similarity functions such as the Gaussian kernel and cosine (2) the number of landmark points that are used in landmark-based scalable methods. Experiments demonstrate the effectiveness of the tuning technique.