Alphabetized by first author listed. Presenting author in bold font.

Bagging Bayesian Networks: Investigating Determinants of Disaster Risk
Kobi Abayomi, Dr. Upmanu Lall, Dr. Andrew Gelman, Columbia University
kaa71@columbia.edu
A common quantification of Risk, of a disaster, under spatial independence, is as a probabilistic product over exposed elements and their vulnerability to the disaster. A non-trivial joint quantification of Risk, from multiple disasters, involves the determination of dependencies between the elements and vulnerabilites. A Bayesian Network (BN) - a Directed Acyclic Graph (DAG) where the joint [probability] distribution is the product of marginal, conditionally independent distributions - can be applied to the problem of divining dependency structure. We investigate BN learning on a composite global disaster dataset of large dimension ($n=15600,k>6$) using the DEAL algorithm - which reduces the NP-complete problem by using a heuristic search with random restarts. The DEAL algorithm is highly sensitive to perturbations in the learning set - to improve accuracy, we apply Bayesian Aggregating (Bagging) across many learned networks.

On the Development of an Inexact Newton Trust Region Interior-Point Algorithm for Large-Scale Nonlinear Programming Problems
Dr. Miguel Argaez, Dr. Leticia Velazquez, Jaime Hernandez Jr., University of Texas-El Paso
jaimeh@utep.edu
We present an algorithm for solving large-scale nonlinear programming problems. We use interior-point methodology, a trust region globalization strategy, and conjugate gradient, with Steihaug's ending conditions, to find a solution of the problem. Some preliminary numerical results are presented.

Compact Routing in the Name Independent Model
Marta Arias, Lenore J. Cowen, Kofi A. Laing, Rajmohan Rajaraman, Orjeta Taka, Tufts University
laing@eecs.tufts.edu
This poster discusses compact routing in the name independent model first introduced by Awerbuch et al. for adaptive routing in dynamic networks. Compact routing is the problem of finding good tradeoffs between the amount of space used to store routing tables and the length of the paths the routing algorithm specifies. For space we focus on the maximum space required per node in a network, and for path lengths we consider the stretch, which is defined as the maximum ratio of the length of a path taken between a pair of nodes, to the length of the optimal path between those two nodes. Our work presents solutions in the name-independent model, referring to our assumption that the graph may not be relabelled in a way that encodes topological information. A compact routing scheme that uses $\tilde{O}(n^{1/2})$-sized local routing tables, $O(\log^2 n)$-sized packet headers, and stretch bounded by 5 is obtained. Alternative schemes reduce the packet header size to $O(\log n)$ at cost of either increasing the stretch to 7, or increasing the table size to $\tilde{O}(n^{2/3})$. For smaller table-size requirements, the ideas in these schemes are generalized to a scheme that uses $O(\log^2 n)$-sized headers, $\tilde{O}(k^2n^{2/k})$-sized tables, and achieves a stretch of $\min\{ 1 + (k-1)(2^{k/2}-2), 16k^2+4k \}$, improving the best previously-known name-independent scheme due to Awerbuch and Peleg.

Fermion Monte Carlo Calculations for the Beryllium atom
Alan Aspuru-Guzik, Malvin H. Kalos, William A. Lester, Jr., University of California, Berkeley
alan@aspuru.com
Most calculatins of atoms and molecules using Monte Carlo methods present a "sign problem" for which the most common solution is to apply the fixed-node approximation. An extensive body of work by several authors has shown that this approximation is very effective, but the fixed-node error cannot be estimated a priori. An exact Monte Carlo method for Fermions has been recently proposed by Kalos and Pederiva(1) and extended to molecular systems by Kalos and Hood. We apply the method to the ground state of the Beryllium atom.

The Effects of Pedagogical Agent Gender and Ethnicity on Student Motivation
Amy L. Baylor, E Shen, Yanghee Kim, Xiaoxia Huang, Florida State University
baylor@coe.fsu.edu
Recent empirical studies have shown that animated pedagogical agents have great potential for serving as effective human-like mentors in computer-based learning environments (e.g., Atkinson, 2002; Baylor, 2002; Moreno, Mayer & Lester, 2001). One critical factor in determining the potential effectiveness in support of learning is the extent to which they are perceived by learners as viable mentors. In two controlled studies we examined the effect of pedagogical agent image, by gender and ethnicity, on undergraduates' perceptions and motivation toward learning.

From one basic face shape, eight agent images differing by ethnicity (Black, White), gender (male, female), and realism (cartoon, realistic) were constructed by a graphic artist. After validating their operational effectiveness, these images were implemented as animated pedagogical agents within the MIMIC (Multiple Intelligent Mentors Instructing Collaboratively) agent-based research environment. Importantly, each agent was implemented within MIMIC with identical animations, scripts, and computer-generated voices, differing *only* by image.

In the first "choice" study, 167 participants (99 White and 68 Black) were presented with all eight agent images simultaneously and were asked to select the agent from which they would like to learn, and then learn from it. Results indicated that African American students were more significantly likely to choose an African American agent (although White students were not more likely to choose a White agent), and females were more likely to choose a cartoon agent. Open-ended responses revealed that Black students were much more likely to choose an agent that they could 'better relate to' in terms of ethnicity and gender. After students actually worked with the agent that they chose, other differences emerged. Across all students, Black agents were perceived to be significantly more interesting and able to keep learner's attention after they were chosen. Female agents were perceived as more enthusiastic, interesting, having a personality, and essential for information to make sense once they were chosen. Realistic agents were perceived as more enthusiastic and making instruction more interesting.

In the second experimental study, 273 participants (163 White and 110 Black) were randomly assigned to one of four agent conditions in a two-factor (agent gender x agent ethnicity) between-subjects design. Results indicated that when students had an agent of the *same* ethnicity, they reported significantly higher satisfaction with learning from the agent and perceived the agent as more motivating and helpful. Interestingly, no matter which agent they received, female students evaluated it overall more positively than males in terms of friendliness, helpfulness, and personal warmth.

A Multi-Agent System to Improve the (RAS) of Large High-Performance Computational Clusters
Nina Berry, Jim Brandt, and Ann Gentile, Sandia National Laboratories; Rose Yao, University of Nebraska, Lincoln
ryao@unlnotes.unl.edu
Large computational clusters rely on different levels of "RAS" - Reliability (infrequency of problems), Availability (usability during failures or maintenance), and Serviceability (ease of maintenance and problem diagnosis). Current methods of ensuring "RAS"; use predefined events that signals a single central management node which invokes the appropriate scripts. This design is very limited in the complexity of the situation it can handle and in its scalability.

The MAS for RAS project is focused on developing software to address RAS problems in large computational clusters by distributing agents with decision-making capabilities on each node in the cluster. Each RAS entity will run on its own service processor, therefore the performance of the node will not be compromised. To simulate this situation, we are currently using Zaurus PDAs acting as "service processers"; to run our software.

The MAS for RAS software has several advantages over existing systems. First, it is a decentralized system, which means the RAS of the entire cluster is no longer dependent on one node. Second, the software agents will be independent and is capable of making local and global decisions. This makes the system quasi self-healing and minimizes the work of the system administrator. For those reasons, our system can handle more complex situations and is easily scalable to a large cluster.

Text-Constrained Speaker Recognition Using Hidden Markov Models
Kofi A. Boakye, University of California, Berkeley
kaboakye@icsi.berkeley.edu
This poster presents a possible application of a text-dependent speaker recognition system within the unconstrained domain of telephone conversation speech, as contained in the Switchboard I corpus, a standard corpus in the speaker recognition community. The system utilizes word-level Hidden Markov Models to generate likelihood scores for key words among the backchannel, filled pause, and discourse marker categories.

Examining Cross-Generational Collaboration in a Visual Programming Environment
Jamika D. Burge, Virginia Polytechnic Institute and State University
jaburge@vt.edu
Our research investigates how community members work together to use computing technology in an effort to understand more about their community and its members. These members can also share their ideas and opinions about how to make their community better. Specifically, we are interested in how elderly community members collaborate with student community members as they work together in a visual programming environment. From our research, we expect to understand community collaboration at two levels. First, we want to study novice users in and across generational pairings. Want hope to understand how senior community members influence the younger student community members, and vice versa. Secondly, we want to investigate roles and how they are reciprocated during programming activities. So, instead of assigning specific roles to the elder-student pairings, we are interested in examining how people fall into their roles naturally.

Implementing an Algorithm to Solve Sequential Testing Procedure
Kalatu Davies, Rice University
kdavies@rice.edu
Cervical pre-cancer is a disease that affects many women across the world and if left untreated it can lead to infertility and cancer of the cervix. Thus it is very important to be able to properly diagnose the disease in the pre-cancerous stage. There is a standard testing sequence that is currently being used for the detection and treatment of pre-cervical cancer. We want to use sequential decision analysis to determine if the current standard of care is optimal and the decision rules for each stage of a disease testing sequence. This optimal decision procedure has been solved in the past using backward induction methods for a two stage, binary procedure. However, this method poses many computational limitations. Thus, the main objective of my research is to develop a method for implementing an algorithm to solve the sequential decision problem in relation to disease screening, specifically for cervical pre-cancer. This algorithm may be applicable to many other disease screening procedures.

Bifurcations and Simulations of Jeffery-Hamel Flows
Jessica Deshler, University of New Mexico
deshler@math.unm.edu
Nearly all industrial machines which involve fluid flow have a point at which the fluid must flow from a small area to a larger area, or more simply, through an expanding channel. For a particular wedge angle and at some critical Reynolds number a bifurcation in the flow occurs and the flow changes from pure outflow to flow with regions of outflow and inflow. Clearly this limits the throughput of the channel and thus the efficiency of any machine with such a design. The more we understand the behavior of these flows, the better able we will be to build efficient machines. Simulation of this radial and two-dimensional flow is done using MPSalsa, a finite element CFD code developed at Sandia National Laboratories, which solves the Navier-Stokes system of equations on the geometry of the wedge. The two dimensional simulations are validated via comparison to experimental data and results from basic numerical codes written to solve the Jeffery-Hamel system of equations. By comparing numerical, computational and experimental data, these results are perhaps among the most complete descriptions of Jeffery-Hamel flows. These results may also have future implications in the determination of the selection mechanism as Jeffery-Hamel flows exhibit multiple solutions.

Bivariate Mean Residual Lifetime Function
Musie Ghebremichael, Javier Rojo, Rice University
musie@stat.rice.edu
In survival analysis the additional lifetime that an object survives past a time is called the residual life function of the object. Mathematically speaking if the lifetime of the object is described by a random variable then the random variable is called the residual life random variable. The quantity is called the mean residual lifetime (mrl) function or the life expectancy at age .

There are numerous situations where the bivariate mrl function is important. Times to death or times to initial contraction of a disease may be of interest for twin studies in humans. The time to deterioration level or the time to reaction of a treatment may be of interest in pairs of lungs, kidneys, breasts, eyes or ears of humans. In reliability, the distribution of the life lengths of a particular pair of components in a system may be of interest. Because of the dependence among the event times, we cannot use the univariate mrl function on each event times in order to assess the aging process. A bivariate mrl function is useful in analyzing the joint distribution of two event times where there is dependence between the event times.

In recent years, though a considerable attention has been paid to the univariate mrl function, relatively little research has been devoted to the analysis of bivariate mrl function. The purpose of my work is to extend and apply the concept of mrl functions to a problem that arise in a bivariate survival analysis.

A solver for the maximum-weight independent set problem
Illya V. Hicks, Jeffrey S. Warren, Texas A&M University
j-warren@tamu.edu
For a weighted simple graph G=(V,E), the maximum-weight independent set (MWIS) problem is that of finding an independent set of vertices such that the sum of the weights of these vertices is maximum among all independent sets of the graph. The problem is famously NP-hard. Yet, because of its numerous applications (e.g., in coding theory and computer vision) and its relationship to other interesting and difficult computational problems (e.g., the minimum coloring problem), it is worthwhile to develop exact algorithms that can solve the problem on small graphs in a reasonable amount of time. Balas and Xue developed a branch-and-bound algorithm for the MWIS problem that solves the MWIS problem quickly on a chordal subgraph, finds a larger subgraph on which the solution is still optimal, and then uses the resulting subgraph to make branching decisions. We, too, find subgraphs for which the MWIS problem is easily solved and use them to make branching decisions. However, we construct a high-weight independent set first and then build the subgraph around it. Our goal is to produce larger subgraphs than do Balas and Xue, thereby producing fewer child nodes at every branching instance. Herein, we present our algorithm and compare its performance to other algorithms for the MWIS problem, with special attention to the Balas-Xue algorithm. We also discuss supplementary upper bound computations for these two algorithms, noting their effect on branch-and-bound tree size and run time.

The Development of a Program to Simulate Contact Mode Atomic Force Microscopy
Divine Kumah
divineknjr@yahoo.com
The atomic force microscope (AFM) is a powerful scanning probe technique used for high-resolution imaging and characterization of nanoscale surfaces of various materials, including silicon. Atomic force microscopy allows the researcher to obtain quantitative information on the surface topography and adhesion activity as well as on the micromechanical properties of the superficial layers of materials. The atomic force microscope probes the surface of a sample with a sharp tip a few microns long and around 100 angstroms in radius. The tip is mounted at the free end of a cantilever that is between 100 and 200 micrometers in length. One major drawback identified in AFM imaging is the dependence of the image's precision on the shape of the probe tip. Artifacts are introduced during AFM imaging as a result of convolution between the tip and the sample. This study aims at providing a simulation to investigate artifacts in Atomic Force Microscopy. A program has been developed to simulate AFM in the contact mode to investigate the effect of tip design on the quality and accuracy of AFM images. Tips of varying dimensions are used in the simulation program to image a sample surface which has features suspected to produce artifacts. The images produced are analyzed and compared to evaluate tip-image convolution. This program shows promise as a tool to help scientists in the measurement and characterization fields, separate true images from artificial images in AFM.

Efficient Gradient Estimation for Motor Control Learning
Gregory Lawrence, Noah Cowan, Stuart Russell, University of California, Berkeley
gregl@cs.berkeley.edu
The task of estimating the gradient of a function in the presence of noise is central to several forms of reinforcement learning, including policy search methods. We present two techniques for reducing gradient estimation errors in the presence of observable {\em input} noise applied to the control signal. The first method extends the idea of a reinforcement baseline by fitting a local model to the response function whose gradient is being estimated; we show how to find the response surface model that minimizes the variance of the gradient estimate, and how to estimate the model from data. The second method improves this further by discounting components of the gradient vector that have high variance. These methods are applied to the problem of motor control learning, where actuator noise has a significant influence on behavior. In particular, we apply the techniques to learn locally optimal controllers for a dart-throwing task using a simulated three-link arm; we demonstrate that the proposed methods significantly improve the response function gradient estimate and, consequently, the learning curve, over existing methods.

Infrastructure for Performance Tuning LAM/MPI Applications
Kathryn Mohror, Karen L. Karavanic, Portland State University
kathryn@cs.pdx.edu
Clusters of workstations are becoming increasingly popular as a low-budget alternative for supercomputing power. In these systems, message-passing is often used to allow the separate nodes to act as a single computing machine. Programmers of such systems face a daunting challenge in understanding the performance bottlenecks of their applications. This is largely due to the vast amount of performance data that is collected, and the time and expertise necessary to use traditional parallel performance tools to analyze that data. Paradyn is a parallel performance tool that addresses these issues by employing dynamic instrumentation to insert the performance measurement instructions into the application and by automatically locating bottlenecks for the programmer. This project implements support for LAM/MPI into Paradyn. LAM/MPI is one of the two most important implementations of the Message Passing Interface (MPI), and also includes several newer MPI features, such as dynamic process creation. As a result of this project, parallel application programmers will be able to use LAM/MPI and have access to detailed performance data while developing their applications. This project will also lay the foundations for future performance tool support for the newer features of MPI.

Reverse Engineering of Genetic and Protein Networks
Edusmildo Orozco , Dr. D. Bollman, Dr. O. Moreno, University of Puerto Rico at Mayagez,
eorozco@cs.uprm.edu
The graphical representation of a network consists of a set of nodes and a set of edges that connect certain pairs of nodes. In a genetic network the nodes represent genes and an edge from node g1 to node g2 represents the idea that a change in the activity of gene g1 changes the activity in gene g2. In a protein network each node represents a protein and an edge from node p1 to node p2 represents the idea that protein p1 interacts with protein p2. Traditionally the networks used for such modeling purposes are Boolean. In such a model, either one gene can affect another or not and either one protein reacts with another or not. We consider a more general type of network in which nodes are represented by variables that vary over finite fields. These networks allow for more realistic models, in which the effect of one gene on another or the rate of a chemical reaction involving two proteins can be measured over a full range of discrete values. Although reverse engineering methods have been applied mostly to genetic methods, it could be advantageous to apply them to protein networks as well. The reverse engineering problem is the problem of determining the network, given experimental data. In this ongoing work we study efficient algorithms for solving the reverse engineering problem for networks over finite fields.

The African American Distributed Multiple Learning Styles System (AADMLSS)
Nicholas Parks, Tanecia K. Simmons, Juan E. Gilbert, Auburn University
gilbert@eng.auburn.edu
Each student has a personal learning style that originates from innate tendencies and environmental experiences. Because cultural groups often share common values, the experiences of children growing up with those values are reflected in their classroom learning behaviors (i.e. cultural learning style). Therefore, a culturally relevant pedagogy is central to the academic success of minority students. The research described in this talk is influenced by the compelling impact of social and cultural issues on academic performance. Accordingly, the African American Distributed Multiple Learning Styles System (AADMLSS) was developed to provide Educators with an easy to use viable alternative, for supplementing their classroom instruction portfolio, with culture specific e-learning tools. AADMLSS embraces the differences in cultural learning styles by providing a culturally sensitive, multi-curriculum, e-learning pedagogical environment, in an effort to enhance a student's overall learning experience and classroom performance. In this presentation, we present the infrastructure and empirical data findings for AADMLSS.

Aerodynamic Simulation of a Falling Paratrooper: Mesh Refinement Techniques for Higher Accuracy in the Computational Boundary Layer
Victor Udoewa, Tayfun Tezduyar, Rice University
udoewa@rice.edu
Our target is to develop computational techniques for studying aerodynamic interactions between multiple objects with emphasis on studying the fluid mechanics and dynamics of an object exiting and separating from an aircraft. The object could be a paratrooper jumping out of a transport aircraft or a package of emergency aid dropped from a cargo plane. These are applications with major practical significance, and what we learn and develop can make a major impact on this technological area. The computational tools we are developing are based on the simultaneous solution of the time-dependent Navier-Stokes equations governing the airflow around the aircraft and the separating object, as well as the equations governing the motion of that object. And the computational challenge is to predict the dynamic behavior and path of the object, so that the separation process is safe and effective. The gravitational and aerodynamic forces acting on the object determine its dynamics and path. We are focusing on more accurate computation of the boundary layer around the aircraft, so that the aerodynamic forces acting on the paratrooper during the period immediately following the exit from the aircraft are calculated more accurately.

A Reduced Basis Method for Molecular Dynamics Simulations
Rachel E. Vincent, Dr. Danny C. Sorensen, Rice University; Dr. B. Montgomery Pettitt, University of Houston
rvincen@caam.rice.edu
Molecular dynamics (MD) simulations allow scientists to computationally determine atomic positions in a molecule over a specified period of time. These simulations are computationally expensive and generally require a large amount of data storage. We expect to reduce computational costs and storage requirements using singular value decomposition (SVD) analysis. In any trajectory, whether generated by traditional dynamics methods, time-averaged refinements, or a reduced basis set method, classical principal component analysis may be used to classify and represent the dominant characteristics of the MD trajectory. Here we augment the classical principal component analysis with an SVD updating scheme. SVD analysis of the computed trajectories will be developed to augment abilities to locate active sites, to identify preferred molecular configurations, and to study periodic behavior. Preliminary results obtained with respect to our reduced basis method provide insight into the relationship between the reduced and standard simulations. Furthermore they suggest that constraints are necessary to insure the integrity of the simulated molecule.

Durable Wide Area Archival Storage in OceanStore
Hakim Weatherspoon, University of California, Berkeley
hweather@cs.berkeley.edu
Traditional archival media are rapidly being replaced with digital repositories. This has generated a need for long-term, digital archival storage. In this poster we describe the architecture of a global-scale, distributed storage infrastructure that is self-repairing and resilient to faults and malicious attacks. This infrastructure employs erasure-coding to enhance durability, coupled with mechanisms for location-independent routing, introspective failure analysis, and automatic repair. The result is archival storage that has the potential to preserve information indefinitely. We present results from the prototype archival layer of OceanStore, currently under construction at Berkeley. The results include expected throughput, latency, and other application usage details. That is, we present what is required to use the archival layer of oceanstore for our everyday storage needs. By building the system we are able to determine the minimal set of requirements to maintain data reliably in the wide area.

Mathematical Modeling Framework for American Option Valuation with Financial Constraints
Donald C. Williams, Rice University
donald@caam.rice.edu
Fundamental to many complex financial derivative securities is the valuation and optimal exercise of options with American-style exercise features. This remains one of the most important and intellectually challenging problems within option pricing theory. This work proposes a direct computational algorithm for solving the American option valuation problem within an optimization framework. The algorithm employs a Newton type constrained nonlinear interior-point optimization method for solving the discretized variational inequality problem that arises. Considerations are made in terms of numerically approximating, in a stable manner, the governing parabolic partial differential equation that can become convection dominated in certain areas of the solution domain. Considerations are also made for incorporating additional economic constraints within the optimization framework. Some example computations are presented for special cases of American-style options with the aim of revealing the general applicability of the constrained optimization pricing methodology.

Optimization of Trajectories to Mars Using Electric Propulsion
Powtache Williams, Rice University
powwow@rice.edu
Although chemical rocket propulsion is widely used in space transportation, large amounts of propellant and vehicle mass limit designs for a human mission to Mars. Electric propulsion, which requires a smaller propellant load while producing greater speed, is an alternative system that can be used for manned interplanetary flight. This research investigates the use of ion electric rockets in previous robotic missions for the design of manned missions.

Specifically, this research modifies the sequential gradient-restoration algorithm (SGRA) to a multiple-subarc sequential gradient restoration algorithm (MSGRA) and includes specifications of ion electric engine (Isp = 3,000 sec, Deep Space 1 Engine) for optimization of maximum payload and minimum time trajectories. Additional studies will be made on the design of hybrid launch vehicles, which includes the use of chemical engines for flight from Earth to low-Earth atmosphere and ion electric engines for interplanetary space flight for both manned and unmanned missions.

Multimodal SQL - Mobile Access to Databases
Dale-Marie Wilson, Auburn University
wilsodc@eng.auburn.edu
The increasing sizes of databases coupled with the decreasing sizes of mobile computing devices introduce increasing restrictions on information display and output. The ballooning sales of these devices and the rising number of Internet users, have led to the onset of pervasive human-centered computing. New advances in speech and multimodal interfaces complement this trend. This poster presents Multimodal SQL, a speech interface that allows users to remotely access their databases using their voices. Multimodal SQL consists of three major components: a speech interface coded using VoiceXML and JavaScript; a relational database; and an interface for the presentation of the results. Multimodal SQL is the first iteration of this software development cycle and its results are presented visually. The cycle is projected to conclude in an interface that accepts spoken queries and returns visual and auditory results. A usability study to determine both the effectiveness and user satisfaction of the first iteration will be presented. The results of this study indicate that Multimodal SQL provides efficient access to databases using voice queries.

Learning rules for the low-dimensional Clifford neural networks
Qing Yi, Bryant York, Portland State University
yiq@cs.pdx.edu
Low-dimensional Clifford algebras include the real numbers, the complex numbers, and the quaternions. Most neural network theory is applied to the learning (or approximation) of real-valued functions of a real variable. In the past decade researchers have begun to study the approximation of functions defined on the complex numbers and the quaternions. A number of theoretical results exist; however, the extension of the universality results from the real numbers to these arbitrary low-dimensional Clifford algebras has met with some resistance. Because the Clifford algebras are generally normed, associative algebras (not necessarily commutative), Clifford analysis poses different problems from real and complex analysis. The learning rules for multilayer perceptrons using the backpropagation algorithm are critically dependent on gradient-descent and the notion of differentiability in the optimization space. In this work, we outline the essential differences between, real, complex, quaternionic, and Clifford analysis as they pertain to the development of effective learning rules for Clifford neural networks. In addition, we demonstrate the effectiveness of these learning rules for a variety of applications.