Skip to main content

Secondary navigation:

Main navigation:

2006 Bachelors Theses

computer science

Author: Daniel J. Scali
Title: Evolving Strategies for the Repeated Prisoner’s Dilemma Game with Genetic Programming: Studying the Effect of Varying Function Sets
Advisor: Sergio Alvarez
This thesis examines the application of genetic programming to evolving strategies for playing an iterated version of the Prisoner’s Dilemma game. The study examines the evolution of strategies for a single population of players pitted against a static environment, as well as the co-evolution of strategies for two distinct subpopulations of players competing against one another. The results indicate that the strategies that can be evolved are strongly influenced by the function set provided during the design process. In co-evolutionary runs in particular, the function set shapes the environment of opponents that an individual strategy is evaluated against. Experimental runs that alter the makeup of the function set facilitate a discussion of how different function sets and environments can lead to diverse strategies with varying levels of performance.

Author: Robert Russo
Title: Evaluating Bayesian Networks as a Recommender System Technology for Motion Pictures
Advisor: Sergio Alvarez
This thesis applies machine learning techniques to a dataset of movies described by collaborative (social) and content attributes in order to create a mixed recommender system for movies. Bayesian networks, two versions of neural networks, decision trees, and simple rule classifiers are compared. It is determined that Bayesian networks and a 1R classifier based on the single best attribute outperform the remaining techniques. An attempt to contrast recommendation quality for content-only and collaborative-only datasets as compared with a dataset described by both content and collaborative attributes yields inconclusive results. Lack of sufficient content information in the current datasets may be the reason for this.

Author: Sergey Weinstein
Title: 3-D Stereoscopic Reconstruction using Structured Light
Advisor: David Martin
Much like humans, computers are able to infer 3-D positions of objects using sight alone. Replacing our eyes with electronic cameras and our brain with soul-less, heartless algorithms, it is possible for a computer to reconstruct a 3-D scene from simple photos. In order to understand the relative positions of objects in the world, we can take two photos of the same scene (from different viewpoints) and look at how positions of objects change from photo to photo. To illustrate this, try extending your arm and holding it in front of your eye. Look at your thumb with only your right eye, then switch eyes. Your thumb seems to jump a great deal between those two view points, whereas the background stays in mostly the same spot. This change of relative positions allows us to understand the 3-D structure of the things we see. The major difficulty with simulating this on a computer is the fact that its very hard to know where an object is from photo to photo (e.g. a computer doesn't know what a thumb is). The common way to identify the 'correspondences' between photos is to look at the image one small area at a time, and then trying to find a similar patch in the other image. This process is very time consuming and can result in inaccurate correspondences. The strategy taken by this thesis involves using structured light (projecting a checker board on to the scene while photographing it.) This strategy allows the computer see the image as a binary collection of pixels rather than a multi-valued grey (or color) image. The computer can then give a unique index to each part of the scene and quickly figure out where things are in both pictures. It can then recover the 3-D structure of the scene and paste the surfaces with the original objects' textures.