Course-Related Projects

Directed Study on Multiagent Sequential Decision Making Under Uncertainty:
Multiagent Coordination Strategies for Improved Search of the Joint Policy Space.
[2005, University of Michigan, Advisor: Ed Durfee]
I analyzed an existing method for stochastic multiagent planning (JESP) and extended it with heuristics for improved quality, based on optimality metrics that I developed.

Robotics Project: Lyapunov-Guided Reinforcement Learning.
[2005, University of Michigan, Course Instructor: Brent Gillespie]
I applied a recent RL algorithm, combining tile-coding function approximation and Lyapunov control theory, to the classic pendulum swing-up control problem, and evaluated its performance trade-offs.

Real-Time Systems Project: Multi-agent Gallery Surveillance.
[2005, University of Michigan, Course Instructor: Kang Shin]
I led a team of two other students and myself in the development, simulation, and evaluation of a robust and reactive surveillance system, whereby autonomous vehicles equipped with potential-field-driven Voronoi-graph-based path planning algorithms patrolled and responded to threats in a coordinated fashion.

Advanced Machine Learning Project: Iterative Rank Refinement.
[2003, Cornell University, Course Instructor: Thorsten Joachims]
Extending a neural net algorithm called RankProp (developed by Caruana et al.), I proposed a new method for learning to rank examples using support vector machines (SVMs), which we compared with several other SVM ranking schemes.

Heuristic Optimization Project: Comparison of Stochastic, Methods for Protein Folding.
[2003, Cornell University, Course Instructor: Bart Selman]
The objective of this project was two fold: to find the minimal-energy conformation of the dipeptide 2-alanine molecule by optimizing a complex cost function, and to compare the applicability and performance of three stochastic methods to this problem. I worked with two other students to implement and apply our own flavors of Simulated Annealing, Tabu Search, and Genetic Algorithms. We derived statistical results comparing the methods and presented what we found to be the optimal configuration of atoms.

Competitive Machine Learning Project: Comparison and Combination of Empirical Methods.
[2002, Cornell University, Course Instructor: Rich Caruana]
This project was a sort of mini-competition between small groups that brought together everything that we learned in CS578. I worked with two other students to perform a statistical comparison of the performance of combinations of Decision Trees, K-Nearest-Neighbor models, and Artificial Neural Networks on a mystery data set. Techniques such as bagging, boosting, model averaging, and feature selection were employed in an attempt to seperately optimize accuracy, RMSE, and ROC Area. And our group arrived at models that tested among the best.

Undergraduate Project: Musical Instrument Recognition.
[2002, Cornell University, Course Instructor: Golan Yona]
I created an automated system to differentiate several musical instruments in recorded samples. The system, which combined methods from machine learning and digital signal processing, was evaluated and shown to be quite successful on test set collected from a free online instrument database.

Click here to go back to my more recent projects