Research Experience for Undergraduates

UTEP Summer Site in Applied Intelligent Systems

Sample Research Projects

The unifying research theme for this REU will be the use of intelligent system techniques, including machine learning, data mining, optimization, and image analysis, to solve relevant data analysis problems in science and engineering fields. Students will be able to choose from the following list, according to their interests and abilities:

Discovering the Patterns of Interaction in Spoken Dialog

Spoken dialog systems today are not as responsive as one would like; indeed, their patterns of interaction are generally perceived as rigid and robotic. Improving this requires a better understanding of what patterns of interaction are normal in human-human interaction. In this project we will develop tools and techniques for both semi- and fully-automated analysis of medium to large dialog data corpora, and measure the value of the discovered patterns for dialog systems.

Generating Test Cases for Pairwise Testing Using Genetic Algorithms

Pairwise testing is a combinatorial testing technique that tests all possible pairs of input values. Although, finding a smallest set of test cases for pairwise testing is NP-complete, pairwise testing is regarded as a reasonable cost-benefit compromise among combinatorial testing methods. The objective of this project is to formulate the problem of finding a pairwise test set as a search problem and apply a genetic algorithm to solve it. A genetic algorithm is a technique that simulates the natural process of evolution and is known to be very effective for finding solutions for problems with a huge search space and complexity. The project will also develop an open-source program that could be used as a framework for generating pairwise test sets using genetic algorithms.

Automated Identification of Foreign Accents in Speech

In many security applications, it is desirable to determine a person's native language from speech in a second language. The first part of this project will consist of developing a small corpus of English spoken by speakers of English as a second language with different native languages (for example, Spanish, Chinese, Arabic, Portuguese, and French). In the second part, various features and learning algorithms will be evaluated to determine the degree to which native languages can be accurately identified automatically and the features (phonemes, prosody, etc.) that allow this identification.

A Survey of Fuzzy Measure Extraction Techniques and Design of a Hybrid Optimization Technique for Multi-criteria Decision Making

At the heart of decision making are very often multiple criteria. It can be challenging to combine them properly to take into account potential dependency between criteria and to make a sound decision. Fuzzy measures combined into a Choquet integral are an option of choice to express dependencies and reach a sound decision. However, fuzzy measures are hard to determine: either the user has to exhaustively describe it (which can be unreasonably long and irrelevant) or the measure can be extracted from some sample data (by solving an optimization problem).

GPU Implementation of Tracking Algorithms

Modern vision-based tracking algorithms usually use a version of the condensation or particle filter algorithm, which approximates the distribution of possible configurations of the objects of interest using a discrete set of hypotheses commonly called particles. The likelihood of each particle has to be evaluated at every tracking cycle, and in complex domains thousands of particles are needed to properly approximate the target probability distribution. The goal of this project is to implement a vision-based tracker using a Graphics Processing Unit (GPU) in order to evaluate hundreds or thousands of particles in parallel, allowing real-time tracking of multiple complex objects.

Analysis of Animal Behavior Using 3D Computer Vision

Behavior analysis of laboratory animals is a tedious process that has to be performed by visual observation. The goal of this project is to develop a software system based on the Kinect 3D image-based sensor to autonomously monitor and analyze the behavior of laboratory rats.

Game Theory for Homeland Security

A common problem in protecting infrastructure and networks against attackers is making decisions about how to allocate limited resources to protect the most important targets. This work focuses on using game-theoretic modeling to understand these decisions, find optimal randomized (unpredictable) policies for resource allocations, and give recommendations to police or other end users. This project will address specific issues in resource allocation for border security.

Agent Technologies for the Smart Grid

The world is undergoing dramatic changes in the ways that energy is generated, distributed, and used. Intelligent agents are a critical technology in managing the increasingly complex and decentralized energy systems that are evolving. These technologies, often called the "smart grid", have the potential to dramatically improve the efficiency and reliability of future energy markets. This project will focus on developing a trading agent to compete in the "Power TAC" game, a new scenario this year that is part of the annual Trading Agent Competition.

Compression of Meteorological Data

Environmental monitoring stations are generation enormous amounts of data that pose significant challenges for transmission and storage. This project deals with the use of machine learning approaches to attain very high compression rates with minimal information loss. This research would build on similar approaches we have applied successfully in the areas of medical image compression and surveillance.

Picking First-Arrival Traveltimes for Seismic Tomography using Machine Learning

Seismic tomography is a technique used by geoscientists to study the upper regions of the Earth's crust. For active tomography experiments, seismic waves are created using signal injectors or explosions and recorded using geophones or seismographs. Then, first-arrival times are read from the recorded signals and used as inputs for tomography software. The reading process is really a measurement process where the typical sources of error can affect the quality of the read travel times, which are referred to as the pick times. It can take a human analyst up to five years of training and practice to consistently record high-quality pick times. The goal of this project is to develop a method to automate this process, enabling a program to {\it learn} the rules that are used to determine pick times given data generated by human experts.

Automated analysis of astronomical images

With the current automated systematic sky surveys, such as the Sloan Digital Sky Survey, astronomers are facing an enormous data overload. To take advantage of all this data, intelligent techniques for automated analysis have to be developed. The goal of this project is to develop methods based on machine learning and computer vision for the automated analysis of multispectral astronomical images in order to enable the identification of stars, galaxies and peculiar objects.

Effects of Resolution-enhancement techniques on face recognition by humans

We propose to investigate the effects of resolution enhancement, occlusion removal, and foveation as aids to improve face recognition from wide-field surveillance imagery by humans and computers. We will analyze strengths and weaknesses of each method and establish guidelines for the design of next generation image processing systems for face recognition in surveillance applications.

Compressive sensing for multispectral imaging

Multispectral imaging is an imaging modality where a stack of images sampling the electromagnetic spectrum are collected simultaneously. When the electromagnetic spectrum is sampled continuously at a high spectral resolution it is called hyperspectral imaging. In hyperspectral imaging there is usually a trade-off between temporal and wave-length resolution. Compressive sensing (CS) is a new technique for reconstructing signals from a samlle number of measurements; in this project we will explore the use of compresive sensing as a way of attaing high temporal and wave-length resolution in hyperspectral imaging.

Back to REU homepage