Abstracts of Previous Seminars

RIACS Seminar #148-A

Date: August 12, 2002

Title: "SSRP Presentations, Group 4"

Speaker(s) & Title(s):

Rene Vidal of U.C. Berkeley: "Bayesian Motion Estimation and Surface Reconstruction"
Kate Mullen of Bard College: "Initialization of Model-Based Clustering Algorithms"
Brent Venable of University of Padova: "Finding Optimal and Pareto Optimal solutions of Simple Temporal Problems with Semi-convex Preference Functions"

RIACS Seminar #148

Date: August 22, 2002

Title: "TrustNEt"

Speaker(s): Dr. N.S. Sridharan

Affiliation(s): TrustNet

Abstract:
Vision: To advance trust in organizations, communities and societies at large. At present TrustNet is a consortium in its formative stage and the founders are engaged in discussions with trusted colleagues to get feedback and advice on the TrustNet vision, scope of activities, and sources of revenue.

Trust Technology for Socio-technical systems. Socio-technical systems can be designed with explicit attention to trust. To do this well, we need to understand the many dimensions of Trust and the dynamics of how trust is eroded or supported by technology. We need to start with a grounding of the core concepts developed about Trust from disparate fields such as psychology, philosophy, social sciences, economics and political science.

Many current systems are designed with a technical focus and hence concentrate on creating systems that can be trusted by people. This is one-sided trust. Complementing this is bilateral trust of people (people to people and system to people). Such thinking may be applied to any number of socio-technical systems (Next generation Internet, Air Transportation, Large scale factory systems, ocean going platforms, hospitals etc). Trust is often confused with reliability, safety, security and privacy protection.

Those of us who were early users of Arpanet remember the pre-boom Internet as a trusted Internet; our intellectual crown jewels all resided on this network - data, models, code and reports. We examine the possibility of rebuilding a Trusted Internet based on sound understanding of how trust is generated, maintained and supported. Implications for designing other socio-technical systems will be explored in this seminar.

Speaker's Bio:
Dr. Sri (pronounced Shree) Sridharan was the chief architect for knowledge management at Intel until 2000. At Intel he was widely known and recognized for his vision and strategy that balances business savvy with pragmatism and people sensitivity. He is a trained storyteller and is the initiator of various storytelling initiatives for Intel. He also participated in Manufacturing Systems Architecture, Enterprise Application Integration, and Strategic Information Systems.

Prior to Intel he spent many years as an academic (Stanford, Rutgers, TU Munich) and managing R&D groups in several industries (BBN, FMC, Intel). His specialty is Artificial Intelligence and his PhD is in Computer Science. His work in AI explored applications in organic chemistry, psychology, genetics, manufacturing and legal reasoning. He was Program Chair of IJCAI-89 and on the editorial board of Artificial Intelligence.

He currently is developing a workshop for corporate executive teams on Leadership, Strategy and Trust. He is working a book on the same subject. He is co-founder of a consortium called TrustNet that brings technology to bear constructively on trust, making sense and listening.

He is actively involved in several initiatives to improve education and community-based leadership in poor regions of the world.

RIACS Seminar #147

Date: August 13, 2002

Title: "Disfluencies in Spontaneous Speech - A Speech Technology Challenge"

Speaker(s): Robert Eklund

Affiliation(s): Telia Research, Sweden

Abstract:
In recent years, automatic speech recognition (ASR) systems have attained accuracy levels on constrained tasks that are sufficient for many commercial purposes. However, for more open-ended speech input, robustness remains an important issue. One important question to be addressed is the processing of disfluencies (DFs), i.e. phenonema like pauses (filled or silent), repetitions, truncated words, repairs, and so on, which occur frequently in spontaneous speech, which is why modelling of DFs is crucial for automatic systems that interact with humans in spoken language. Disfluencies have been studied both within and across languages, as well as within and across domains and modalities. This talks has two different parts. First, a general introduction to DFs is given, including a typology of different, as well as different implications for different fields of research, as well as an introduction to different data collection methods, and various ways to transcribe and annotate the data collected. Second, disfluency research at Telia Research AB, Sweden, is described in some detail. Four travel dialogue corpora are described, with four different settings: Human--"Machine"--Human (Wizard-of-Oz); Human--"Machine" (Wizard-of-Oz); authentic Human--Human and authentic Human--Machine. Results on five different kinds of disfluencies are presented: filled and unfilled pauses, prolonged segments, truncations and explicit editing terms. Also, a brief discussion on cross-language aspects of DFs is given, with examples from American English, Swedish and Tok Pisin.

RIACS Seminar #146

Date: August 8, 2002

Title: "SSRP Presentations, Group 3"

Speaker(s) & Title(s):

Cleidson de Souza of U.C Irvine: "A Field Study of Collaborative Software Development Teams"
Sarfraz Kurshid of MIT: "Symbolic Execution for Concrete Test Case Generation"
Frank Hutter of Darmstadt University Of Technology: "Model-based Fault Diagnosis and State Estimation for Planetary Rovers"
Alex Groce of Carnegie Mellon University: "What Went Wrong?"

RIACS Seminar #145

Date: August 1, 2002

Title: "SSRP Presentations, Group 2"

Speaker(s) & Title(s):

Laurentiu Leustean of University of Bucharest: "Certifying Kalman Filters"
Flavio Lerda of Carnegie Mellon University: "Towards Model Checking of C Programs"
Dan Bohus of Carnegie Mellon University: "Spoken Dialog Management for an Astronaut's Procedure Assistant"
Jamieson Cobleigh of University of Massachusetts: "Assumption Generation for Compositional Model Checking?"

RIACS Seminar #144

Date: August 1, 2002

Title: "SSRP Presentations, Group 1"

Speaker(s) & Title(s):

Max Horstmann of University of Massachusetts: "The Markov Decision Process Approach on Planetary Rover Contingency Planning"
Judah De Paula of University of Texas: "Partitioning Algorithm for Data Tracking"
Nate Blaylock of University of Rochester: "Integrating Planning and Execution for a Dialogue Interface with Autonomous Agents"
Ellen Campana of University of Rochester: "Eye Tracking Technology in Speech Recognition Interfaces"

RIACS Seminar #143

Date: July 18, 2002

Title: "An Integrated Approach to Fault Detection and Isolation in Complex Hybrid Systems"

Speaker(s): Dr. Gautam Biswas

Affiliation(s): Vanderbilt University

Abstract:
The need for reliability and robustness in present day systems requires that they possess the capability for accommodating faults in the controlled plant. Fault accommodation requires tight integration of online fault detection, isolation, and identification with the system control loop. This thesis presents a model-based approach to online fault detection, isolation, and identification in complex systems. The plant models for such systems are necessarily hybrid, i.e., their behavior evolution combines continuous operating regions (modes) interspersed with discontinuous changes that model mode transitions (the result is a discrete change in the continuous model). Hybrid models form the natural representation for embedded systems in avionics, automotive, and robotics domains. The wide applicability of hybrid systems has inspired a great deal of research from both control theory and theoretical computer science.
A model-based approach to fault detection and isolation of hybrid systems presents an interesting set of challenges that mostly revolve around interactions of the continuous and discrete components of the system. The tracking of the system behavior evolution has to be performed across modes of operation. This requires continuous tracking, identification of discrete changes, and updating the model and state after a discrete change. The fault isolation has to reason across modes of operation to identify hypotheses that can explain all deviant observations. This may involve rolling back in the mode space to generate hypotheses and then rolling forward in the mode space to catch-up to the current system mode of operation. For fault accommodation, the quantitative value of the fault has to be determined so that appropriate corrective action may be taken. We present an integrated diagnosis architecture that tracks the hybrid system, and detects, isolated and identifies the fault. We use hybrid bond graphs as a comprehensive modeling framework from which models for the individual components of our diagnosis architecture are derived. We use these models to develop diagnosis algorithms that combine hybrid behavior tracking with mode detection and combined qualitative-quantitative reasoning techniques in the continuous domain. The effectiveness of the approach is demonstrated for a realistic example: the fuel transfer system of aircraft. These techniques are now being applied to a NASA application, the Water Recovery System (WRS) of the Bio-Plex system for long-term manned missions.

Speaker's Bio:
Gautam Biswas is an Associate Professor of Computer Science and Engineering, and Management of Technology at Vanderbilt University and a faculty associate at the Institute for Software Integrated Systems (ISIS) in the School of Engineering. He has a Ph.D. degree in Computer Science from Michigan State University in E. Lansing, MI. Prof. Biswas conducts research in Intelligent Systems with primary interests in hybrid modeling and analysis of complex embedded systems, and their applications to diagnosis and fault-adaptive control. As part of this work, he is working on fault-adaptive control of fuel transfer systems for aircraft, and the water recovery system of the Bio-Plex system. He is also initiating new projects in distributed monitoring and condition-based maintenance. In other research projects, he is also involved in developing simulation-based environments for learning and instruction, decision-theoretic planning and scheduling techniques for intelligent manufacturing systems, and Hidden Markov Model techniques for clustering of temporal data sequences. His research is currently supported by funding from NASA, DARPA, and the NSF.

Dr. Biswas has served on the Program Committee of a number of conferences. He was chair of the 1997 IJCAI Workshop on Engineering Problems for Qualitative Reasoning, co-chair of the 1996 Principles of Diagnosis Workshop, the 1999 AAAI Spring Symposium on Hybrid Systems and AI, the 2001 Workshop on Qualitative Reasoning, Senior Program committee for AAAI-97 and AAAI-98, and Technical Committee co-chair for the 2000 IEEE SMC conference. He is currently an Associate Editor for the IEEE Transactions on Systems, Man, and Cybernetics and the International Journal of Applied Intelligence. He is also a guest editor of a IEEE Transactions on Systems, Man, and Cybernetics Part B special issue on "Diagnosis of Complex Systems: Bridging the methodologies of the FDI and DX Communities." Currently, Dr. Biswas is a Senior member of the IEEE Computer Society, ACM, AAAI, and the Sigma Xi Research Society.

Contact Information:
Gautam Biswas
Dept. of Electrical Engineering and Computer Science
Box 1824, Station B
Vanderbilt University

RIACS Seminar #142

Date: July 18, 2002

Title: "Explaining Drugs to a Computer: Building NDF-RT - A Reference Terminology for Medications"

Speaker(s): Mark Tuttle

Affiliation(s): Apelon

Abstract:
Drugs are dangerous, expensive and revolutionizing healthcare. Computers need to help us reduce errors, improve quality and manage costs, but they cannot do with this productively without an ability to "interoperate by meaning." Specifically, if a patient's chronic care medication list is in one computer, and his or her acute care medications are being entered into another, the latter computer's decision support can act on both lists only if there is sufficient interoperation by meaning. At present, the best way for two computers to agree that they "understand drugs" is for them both to use the same Reference Model. The VHA (Veterans Health Administration) and Apelon are building such a Model. Beginning with NDF (National Drug File), the current VHA Formulary, we are creating an ingredient-centric, semantic, i.e., computer-understandable, definition for each drug based on chemical structure, mechanism of action, physiologic effect, pharmaco-kinetics/dynamics, and therapeutic intent. The Model allows the more than 80K orderable drugs at VHA to "inherit" the properties of their active ingredients, except when the particular "dose form" dictates otherwise. The initial version of NDF-RT has been computed from NDF and the UMLS Metathesaurus. It can be browsed from the Web. Human editors are reviewing the next version. Of potential relevance to Astro-Biology and Bio-Terrorism is the instantiation of this model on all "biologically active" substances. Attendees at a recent meeting on "Drug Informatics" agreed that the relevant portion of the schema for a semantic definition of medications could and should be applied to all other biologically active substances.

Speaker's Bio:
Mark S. Tuttle is Vice President of Strategy at Apelon (www.apelon.com): Currently, Mark directs a number of efforts to create, deploy, maintain, and use biomedical Reference Terminologies. Mark is a co-founder of Lexical Technology - which later merged with Ontyx to become Apelon - and he proposed and then led the initial development of the NLM UMLS (Unified Medical Language System) Metathesaurus. Prior to founding Lexical, he helped the University of California, San Francisco win and execute a first round UMLS award, while teaching computer science at University of California, Berkeley. He studied computer science, applied mathematics, and information theory at Harvard University and Dartmouth's Thayer School of Engineering. Mark was elected to be a Fellow of the American College of Medical Informatics (FACMI) in 1993.

RIACS Seminar #141

Date: July 3, 2002

Title: "Developing Automated Spoken Language Dialogue Systems"

Speaker(s): Dr. Dominique Estival

Affiliation(s): Defence Science & Technology Organisation, Edinburgh, AUSTRALIA

Abstract:
The main focus of my talk will be the work of the NLP group at Syrinx Speech Systems, and in particular I will describe Sylan, the automated dialogue system we developed. Sylan was fully integrated with the Sycon speech recogniser into a platform combining speech recognition, natural language processing, dialogue management, telephony and database integration and permitting the deployment of natural lan guage dialogue systems in automated call centres. The aim was to produce a framework to build applications where the structure of the dialogue can be less constrained than current commercial directed systems, and which allow users to input more natural, multi-token utterances which are interpreted and processed in several stages. I will first describe the architecture of Sylan, which, being modular, allowed us to build a system with domain-independent components reusable from application to application. I will then present those components from the point of view of application developers, describing the data structures used by the system and the utilities to build them, and drawing examples from the two prototypes we developed. I will conclude by discussing the constraints imposed on the development of commercial systems, and by drawing the lessons learned along the way to point to further research directions.

Speaker's Bio:
Human Systems Integration Group
Information Technology Division
Defence Science & Technology Organisation
PO Box 1500
Edinburgh SA 5111
AUSTRALIA
Phone: +61 (08) 8259 6485 Fax: +61 (08) 8259 5589
Email: dominique.estival@dsto.defence.gov.au

RIACS Seminar #140

Date: July 3, 2002

Title: "On Explaining the Degree of Error Reduction by Combining Classifiers"

Speaker(s): Dr. Kamal Ali

Affiliation(s): Vividence

Abstract:
In the past decade, one of the most interesting developments in Machine Learning has been that combining classifiers in various ways produces ensembles that have lower classification error rates. Looking one step further, I will present results that explain why combining classifiers to produce more accurate classifications helps in some problems much more so than in other problems. The work examines how the noise inherent in the problem limits the lift due to ensembles, how the dimensionality of the problem affects error reduction and the effect that diversity of classifiers has on ensemble error rate. I will also briefly present results from two applications. The first involves recognizing mineral signatures from infrared reflectance spectra using learning expert systems. The second application researched the efficacy of active learning to speed-up human labelling of satellite images and compared a naive Bayes classifier to within- and cross-human labeling variance.

Speaker's Bio:
Dr. Kamal Ali received his Ph.D. in Computer Science from the University of California, Irvine, in December 1996. His dissertation was on learning probabilistic first-order theories but his later work was on learning classifier ensembles. He has published numerous papers in Data Mining and Machine Learning conferences, reviewed for ICML, KDD, PAMI and MLJ. He is the recipient of a NSF scholarship, UC Regents fellowship and the Horner Exhibition prize in Maths at Sydney University. After graduation, at IBM Almaden, he published papers and served as a data-mining consultant. At TiVo he led the team to implement collaborative filtering for TV recommendation used in 400,000 homes. At Vividence he developed a divide-and-conquer system for clustering clickstreams and developed a system for clustering free text. His current research interests include combining classifiers, developing learning algorithms for image analysis, Bayesian statistics, web-mining and text-mining. He is interested in getting involved in mining of scientific datasets and applying learning to space images.

RIACS Seminar #139

Date: June 27, 2002

Title: "Human-Computer Interaction and Augmented Reality Research at Rockwell Scientific"

Speaker(s): Dr. Venkataram Sundareswaran

Affiliation(s): Rockwell Scientific Company

Abstract:
Human Computer Interaction (HCI) work at Rockwell Scientific Company (RSC) includes basic research in developing new HCI tools and integration of multiple modalities to create intuitive interfaces. Basic research includes Speech Recognition in noisy environments and Augmented Reality (AR). In multimodal interaction, research is focused on developing "point and speak" and "look and speak" interfaces, and in developing an Integrated Displays Testbed that includes various levels of displays (handheld, tablet, large screen) and interaction among them. The talk will begin with a broad introduction to the HCI work at RSC and a video presentation that summarizes this work. The second part of the talk will be focused on Augmented Reality technology, with an introduction to AR followed by recent results achieved by the HCI team at RSC. In particular, an approach to solve the "registration" problem, namely that of aligning real world view with virtual graphical objects, based on the technique of visual servoing, will be presented.

Speaker's Bio:
Dr. Sundareswaran leads a team of eight HCI researchers at RSC's Information Sciences division. In this role, he manages Government and corporate research projects. His background is in Computer Vision and Graphics, and has authored numerous publications in these areas. He graduated from NYU in 1992, and after a brief stint as a post-doctoral fellow at INRIA (France), he spent three years at Boston University in Neural Network modeling of visual motion perception. He joined RSC in 1996.

RIACS Seminar #138

Date: June 13, 2002

Title: "PCMon - A Network Monitoring Tool"

Speaker(s): Jerry Toung

Affiliation(s): RIACS, NASA Ames Research Center

Abstract:
PCMon is a network measurement and monitoring tool that is being developed and deployed on the NASA Research and Education Network (NREN) testbed, a research testbed used for prototyping emerging networking technologies. A new tool was required to enable evaluation of the effectiveness of these new technologies. For example, Quality of Service (QoS) is an emerging technology that is designed to provide different levels of service for different traffic flows, a capability that enables efficient sharing of network resources among multiple users while providing preferential treatment to selected applications when network resources become scarce. Traditional network monitoring and measurement tools capture statistics on aggregate traffic only, rather than individual traffic flows, and hence would be unable to verify that "preferential treatment" is actually delivered as promised by the QoS mechanism.

During this presentation, I will to talk about how PCMon has evolved as a functional tool and how it has already been used to monitor QoS requirements on some NASA's applications. I will also do a live demonstrtation of its capabilities and close with few notes on future directions.

Speaker's Bio:
I hold a Master Degree in Electrical Engineering and specialize in network software engineering. I have been with RIACS at the NASA Research and Education Network since January 2000. My interests include network programming and device drivers development for FeeBSD.

RIACS Seminar #137

Date: June 7, 2002

Title: "Dialog Systems that Learn"

Speaker(s): Mohammed El-Beltagy

Affiliation(s): BiosGroup, Inc., in Santa Fe, New Mexico

Abstract:
In many design optimization problems, the designer is faced with the dilemma of how to simulate the problem at hand using a number of different models. Some models maybe quite elaborate in their representation of the problem and hence tend to be computationally expensive. Other models may be far less elaborate and hence computationally cheaper. The computationally cheap models tend to be less accurate than the expensive ones. The designer uses his/her experience, and understanding of the problem domain to switch between different models. S/He goes through a few iterations till a satisfactory design is found. Designs created in such a fashion are not necessarily optimal and they could be improved upon, given more design iterations and an adequate search technique. It is hence important to develop techniques that make maximal use of the many models available within a limited computational budget. Conducting search on such an environment where there are multiple models for evaluating fitness is what is meant by the term Multilevel optimization (MLO). In this presentation, an exposition will made of the issues to be considered when carrying out multilevel optimization. This will be followed by a comparison of how various optimization algorithms perform on a multilevel problem using three simple model selection strategies. Having established that evolutionary inspired search methods work well in such an environment, a topological mapping based model selection approach is then presented. Finally, Gaussian processes based metamodeling and model fusion approaches are explored. We will show that there are significant gains to be made in the synthesis between optimization and machine learning techniques for MLO.

Speaker's Bio:
Mohammed El-Beltagy received the B.Sc degree in Mechanical Engineering from the American University in Cairo, Cairo, Egypt, in 1994, the M.Sc. degree in Mechatronics from Lancaster University, Lancaster, U.K., in 1996, and the Ph.D. degree in Mechanical Engineering from the University of Southampton, Southampton, U.K., in 2000. His PhD. was on multilevel optimization (MLO), where the goal was to optimize an engineering design, the performance of which can be evaluated by a number of simulation models that a tradeoffs between computational speed and accuracy. During the course of his PhD. work, he devised a number of successful strategies to MLO that involved a tight coupling between machine learning and optimization. His work at Southampton was funded by the Engineering and Physical Sciences Research Council (EPSRC), and British Aerospace plc. He is currently a Senior Scientist at BiosGroup, Inc., in Santa Fe, NM. Since joining BiosGroup, he has generated IP around automated financial trading and high dimensional matching, lead a project that devised a new algorithm for the optimization of trading networks for a natural gas pipeline operator. In an extended distributed control problem, Mohammed developed methodologies for self-organizing, self-healing data networks for a large mass storage software provider. His current research interests include search theory, and utilizing machine learning techniques for intelligent optimization and search landscape feature elicitation.

RIACS Seminar #136

Date: June 6, 2002

Title: "Easy-to-Use, Integrated Data Mining"

Speaker(s): David Kil

Affiliation(s): Rockwell Scientific Company

Abstract:
Data mining is considered as an esoteric, domain-dependent art. In this talk, I will explore several key issues in designing an easy-to-use, integrated, and scalable data-mining tool that can be implemented in parallel computers with coarse granularity. The first issue deals with metadata creation by marrying signal processing to data mining in the context of preprocessing, which is the most important step in data analysis. Next, I will discuss automatic database exploration to find entity relations and discover meaningful relationships autonomously. Third, I will investigate the components of an intelligent data-mining expert engine that can guide novice users through a myriad of seemingly complex steps in data mining. Finally, I will demonstrate how these salient concepts can be applied to high-throughput gene-chip image analysis, mine countermeasure, and diagnostic applications. In summary, we will together explore what it will take to change the perception of data mining from an abstract art to a well-understood science.

Speaker's Bio:
David Kil received his BS in EE and Chemistry at the University of Illinois at Urbana-Champaign (Bronze Tablet), MSEE from the Polytechnic University of New York, and MBA from Arizona State University. His primary research interest is in integrated data analysis that combines salient concepts from signal processing, image understanding, data mining, and fusion to extract the maximum amount of useful information from raw data. He has managed over ten 6.1 and 6.2 research programs, dealing with information extraction and knowledge discovery. He is currently leading a data-mining effort at Rockwell Scientific and consults for two biotech companies (Arcturus and Gene Networks) in high-throughput gene-chip image and tissue analyses. He has published over 25 papers and one research monograph titled Pattern Recognition and Prediction with Applications to Signal Characterization by Springer-Verlag.

RIACS Seminar #135

Date: June 5, 2002

Title: "Discovering Regimes in Time Series"

Speaker(s): Dr. Ashok Srivastava

Affiliation(s): Blue Martini Analytic Services

Abstract:
Many real-world time series are multistationary, meaning that the dynamics of a data generating process switches its mode of behavior. These modes of behavior are sometimes referred to as "regimes," and could manifest as a shift in the mean, variance, or some other statistic. Two key problems that arise in the analysis and prediction of such systems are:

· To identify the time at which a system undergoes a mode change, and
· To model the underlying dynamics of the system after the mode change.

In an attempt to address these issues, we will discuss a model that builds on two model classes: mixture models and thermodynamic clustering-models. Using a maximum-entropy framework, we will derive association probabilities, a quantity which represents the probability of associating an input-output pair to a local model. Unlike standard mixture models, these association probabilities are parametrized by a scale parameter that allows us to sweep through multiple time series segmentations. The different segmentations that arise can shed light onto the two problems described above. Applications of such technology arise in a diverse set of fields, including engineering and aeronautical systems, economics, and physiology. We will describe the performance of this model on synthetic and real-world time series.

Speaker's Bio:
Ashok N. Srivastava, Ph.D. is currently Senior Director of Blue Martini Analytic Services and has fourteen years of experience in research and development in the fields of data analysis and machine learning, data mining, signal processing, and applied physics. Some recent research activities include the development of patent pending algorithms for understanding nearest neighbor predictions, methods of clickstream analysis and visualization, creation of new methods to profile populations based on naive Bayes techniques, and the deployment of research results in a variety of industries and government.

RIACS Seminar #134

Date: May 31, 2002

Title: "Information-Theoretic Clustering of Probability Distributions for Semantic Representation"

Speaker(s): Dr. Richard Rohwer

Affiliation(s): HNC Software

Abstract:
This presentation will begin with a brief overview of statistical pattern recognition research at HNC, stressing my particular interests, and then focus in on a specific body of work: a system for clustering probability distributions. This work is interesting for its motivation and for its technical components. The motivation involves the use of probability distributions to embody a very general concept of semantics. The algorithm simultaneously aggregates 2 categories of values into bins with minimal loss of statistically significant mutual information between the two. The number of bins is self-adjusting, trading statistical significance against information preservation. The method is formulated in the multinomial-Dirichlet framework, using the Evidence approximation to adjust the hyperparameters, after verifying the validity of the approximation in this setting. An adaptive simulated annealing procedure is used for training.

Speaker's Bio:
Richard Rohwer graduated with a BS in Physics from Stanford in 1978 and a PhD in Physics from the University of Texas at Austin in 1985, with a thesis on quantum cosmology. Following a 1-year postdoc at the University of Newcastle upon Tyne in England, Dr. Rohwer moved into Neural Networks, working at the Centre for Speech Technology Research of Edinburgh University until 1991. He joined the faculty of Aston University in Birmingham, where he was instrumental in setting up their then new Neural Computing Research Group. In 1996, he moved to Prediction Company in Santa Fe, NM, and on to HNC Software in San Diego in 1997, where he is now a Principal Scientist in the Advanced Technology Solutions Department. Dr. Rohwer currently leads grant-funded research involving diverse applications of statistical pattern recognition, including topics in natural language, machine vision, and bioinformatics.

RIACS Seminar #133

Date: May 30, 2002

Title: "Polychotomic Encoding: A Better Quasi-Optimal Bit-Vector Encoding of Tree Hierarchies"

Speaker(s): Dr. Robert Filman

Affiliation(s): RIACS, NASA Ames Research Center

Abstract:
Polychotomic Encoding is an algorithm for producing bit vector encodings of trees. Polychotomic Encoding is an extension of the Dichotomic Encoding algorithm of Raynaud and Thierry. Polychotomic and Dichotomic Encodings are both examples of hierarchical encoding algorithms, where each node in the tree is given a "gene"---a subset of the integers {1,...,n}. The encoding of each node is then the union of that node's gene with the genes of its ancestors. Reachability in the tree can then be determined by subset testing on the encodings.

Dichotomic Encoding restructures the given tree into a binary tree, and then assigns two bit, incompatible (chotomic) genes to each of the two children of a node. Polychotomic Encoding substitutes a multibit encoding for the children of a node when the restructuring operation of Dichotomic Encoding would produce a new heaviest child (child requiring the most bits to represent a tree of its children) for that node. We prove that Polychotomic Encoding never produces an encoding using more bits than Dichotomic Encoding. Experimentally, Polychotomic Encoding produces a space savings of up to 15% on examples of naturally occurring hierarchies, and 25% on trees in the randomly generated test set.

Speaker's Bio:
Robert Filman is a Senior Scientist at the Research Institute for Advanced Computer Science (RIACS) at NASA Ames Research Center, working on frameworks for developing distributed applications. Prior to coming to NASA in May 1999, Dr. Filman worked in the research groups of Lockheed Martin, IntelliCorp and Hewlett-Packard, and on the faculty of the Computer Science Department at Indiana University, Bloomington. He is Associate Editor-in-Chief of IEEE Internet Computing and is on the editorial boards of the Journal of Software Maintenance and Evolution and the International Journal of Artificial Intelligence Tools. Dr. Filman received his B. S. (Mathematics), and M.S. and Ph. D. (Computer Science) from Stanford University.

RIACS Seminar #132

Date: May 16, 2002

Title: "Image Processing with Complex Wavelets"

Speaker(s): Nick Kingsbury

Affiliation(s): Reader in Signal Processing, University of Cambridge, UK

Abstract:
This talk will describe the Dual Tree Complex Wavelet Transform (DT CWT), which is a form of discrete wavelet transform, which generates complex coefficients by using a dual tree of wavelet filters to obtain their real and imaginary parts. This introduces limited redundancy (2^m:1 for m-dimensional signals) and allows the transform to provide approximate shift invariance and directionally selective filters (properties lacking in the traditional wavelet transform) while preserving the usual properties of perfect reconstruction and computational efficiency with good well-balanced frequency responses. We will then describe briefly why the DT CWT is particularly suitable for images and other multi-dimensional signals, and discuss some applications of the transform that take advantage of its unique properties. In particular we will consider its application to denoising, deconvolution, texture analysis, segmentation, content-based retrieval and watermarking.

Speaker's Bio:
Nick Kingsbury received the honours degree in 1970 and the Ph.D. degree in 1974, both in electrical engineering, from the University of Cambridge. He is a member of the IEEE. From 1973 to 1983 he was a Design Engineer and subsequently a Group Leader with Marconi Space and Defence Systems, Portsmouth, England, specializing in digital signal processing and coding, as applied to speech coders, spread spectrum satcomms, and advanced radio systems. Since 1983 he has been a Lecturer in Communications Systems and Image Processing at the University of Cambridge and a Fellow of Trinity College, Cambridge. He was appointed to a Readership in Signal Processing in 2000. His current research interests include image compression, error-robust source coding techniques, and image analysis and enhancement techniques, particularly those based on wavelet decompositions. He is especially interested in the application of complex wavelets to images and 3-D datasets.

ngk@eng.cam.ac.uk
http://www.eng.cam.ac.uk/~ngk

RIACS Seminar #131

Date: May 10, 2002

Title: "Optimization Over Sequence Spaces"

Speaker(s): Dr. William Macready

Affiliation(s): VP Science at BiosGroup

Abstract:
Landscapes are functions defined over graphs and have proven themselves to be a useful tool for the design and analysis of local search algorithms for combinatorial optimization. Recent advances in kernel methods from machine learning offer the promise of learning landscapes thereby allowing learning to greatly improve optimization. I will discuss beginning steps in this direction for optimization problems defined over sequence spaces. Optimization over sequence spaces includes many important problems from physics (e.g. spin glasses) and biology (e.g. protein folding)

Speaker's Bio:
Dr. William Macready is currently VP Science at BiosGroup, a consulting company applying insights from complexity science, simulation, machine learning, and optimization to solve difficult business problems. William's interests center on the probabilistic inference for the design of efficient optimization and machine learning algorithms. Before joining BiosGroup William worked at the Santa Fe Institute for complex systems as a postdoctoral fellow and at IBM research as a scientist. He has published on the theory of landscapes, molecular evolution, adaptive organizations, economics, optimization, machine learning, and methods of quantifying complexity.

RIACS Seminar #130

Date: May 6, 2002

Title: "Dialog Systems that Learn"

Speaker(s): Alexander Rudnicky

Affiliation(s): School of Computer Science, Carnegie Mellon University

Abstract:
Descriptions of dialog systems commonly focus on core components such as dialog management and understanding. Yet to develop a working system capable of non-trivial operation, a great deal of effort is also spent on defining and implementing the interface between the dialog system and the domain (or "back-end"). Often this aspect comes to dominate the total effort devoted to implementation. We have begun to explore how corpus-based learning techniques can be brought to bear on this problem. This talk will define the problem and will describe our current work on the discovery of operative concepts from transcripts of goal-directed human-human conversation.  

RIACS Seminar #129

Date: May 2, 2002

Title: "Applying Explanation Based Learning to Speech Recognition"

Speaker(s): Dr. Manny Rayner

Affiliation(s): Netdecisions Technology Centre, Cambridge, England

Abstract:
Accurate recognition is essential for speech-enabled command and control tasks. The required accuracy is made possible by using a suitable grammar for constraining the space of recognition hypotheses. For complex applications, construction of this grammar represents a major investment of effort. There is consequently a strong motivation to develop general grammars that can easily be reused between applications. Experience shows however that such general grammars tend to have serious problems with both efficiency and scalability.

In this talk, we will show how it is possible to use Explanation Based Learning (EBL) to specialize a general grammar coded in a high-level, logic-based formalism against a corpus of domain-specific examples. A series of experiments, carried out on the RIALIST group spoken interface to a simulated version of the Personal Satellite Assistant, suggest that specialized grammars have significantly better run-time properties and strikingly better scalability than the general grammars from which they were derived. In particular, the relationship between the number of rules in a grammar and the time required to compile it into an executable form appears to be roughly quadratic for an EBL specialized grammar, compared to roughly exponential for a general grammar.

(Joint work with Beth Ann Hockey and John Dowding, RIACS)

Speaker's Bio:
Manny Rayner is Senior Architect for Voice System Design at the Netdecisions Cambridge Technology Center, where his work focusses on development of tools for implementation of commercial voice-enabled systems. Prior to joining Netdecisions, he headed the RIACS Spoken Language group (RIALIST). He also worked for eight years at SRI International, where he was project lead on the Spoken Language Translator, one of the world's first large-scale automatic speech translation projects. He has over fifty publications, including a book on the Spoken Language Translator which was published last year by Cambridge University Press. He has a B.A. in Mathematics from the University of Cambridge and a Ph.D. in Computer Science from Stockholm University.

RIACS Seminar #128

Date: May 2, 2002

Title: "The Use of Innovative Evolutionary Computation to Solve Intractable Problems"

Speaker(s): Dr. Yuval Davidor

Affiliation(s): Schema

Abstract:
Schema is one of the first and few commercial companies in which its business and core technology was founded on evolutionary computation (EC). Founded in 1994, the Schema group's main purpose was to solve intractable problems in science, industry and commerce. During this time the group addressed various projects such as static and dynamic missile balancing, warehouse shelf space management, container stowage optimization, and NMR field optimization. In 1997 Schema expanded into the wireless industry and introduced its Falcom platform for spectrum management. To date, about 30% of North American cellular subscribers obtain wireless services that have been optimized for capacity and quality of service by the Falcom platform. Based on Schema's optimization technologies that employ network modeling and simulation capabilities, the platform eliminates the tedious trial and error processes required for manually configuring a network for optimal performance. Schema's solutions are deployed and benchmarked by leading wireless operators worldwide, including Verizon Wireless, Cingular Wireless, U.S. Cellular, and BellSouth International.

The presentation will briefly cover past EC applications that the group has developed and will mainly focus on the current Falcom product for optimal frequency allocation and spectrum management.

Speaker's Bio:
Dr. Yuval Davidor is the founder of Schema. Previously he was a scientist in the department of computer science at the Weizmann Institute and was a member of the School for Advanced Studies of the Hebrew University. Davidor has authored and co-authored books, academic publications, and several patents in the field of evolutionary computation and optimization. Since 1991, Davidor has devoted his career to the practical application of optimization technologies to solve intractable problems ranging from defense applications to logistics and telecom. Davidor has a bachelor's degree in Engineering from Tel-Aviv University and a doctorate in artificial intelligence from Imperial College, the University of London.

RIACS Seminar #127

Date: April 23, 2002

Title: "Towards A Generic Spoken Dialogue System"

Speaker(s): James Allen

Affiliation(s): University of Rochester

Abstract:
While there is great interest and activity in building spoken dialogue systems today, most applications involved very limited domains that require no significant reasoning. Our goal is to design and build systems that approach human performance in conversational interaction in domains that require significant reasoning. We limit our study to "Practical dialogues": dialogues in which the conversants are cooperatively pursuing specific goals or tasks. These include planning (e.g., designing a kitchen), information retrieval (e.g., finding out the weather in New York), customer service (e.g., booking an airline flight), advice-giving (e.g., helping assemble some modular furniture) or crisis management (e.g., a 911 center assistant). In fact, our belief is that the class of practical dialogues includes most anything about which people might want to interact with a computer.

While each of these different genres of tasks require significantly different reasoning components and have different structures, we believe that we can develop an generic model of practical dialogue systems that enables us to build domain-independent components that can relatively easily be adapted to different domains. I will describe our work so far and illustrate with examples from some systems we have built over the past five years.

Speaker's Bio:
James Allen received his Ph.D. in 1979 from the University of Toronto. He currently holds the position of Professor (87-present) and holds the John H. Dessauer Chair at the University of Rochester (92-present); formerly he was Department Chair (87-90); Associate Professor (84-87), and Assistant Professor (79-84). He served as Editor-in-Chief, Computational Linguistics (83-93) and was a Presidential Young Investigator (84-89). Dr. Allen is the author of Natural Language Understanding, and Reasoning About Plans, and co-editor of Readings in Planning. Dr. Allen is a Fellow of the AAAI.

James Allen's research interests lie at the intersection of language and reasoning, and span a range of issues including natural language understanding, dialogue systems, knowledge representation, common-sense reasoning and planning. In the last five years, he has been focusing on designing and building end-to-end spoken dialogue systems that require and exploit common-sense reasoning to collaborate with the user in problem solving. The Rochester Intelligent Planning System (TRIPS) is a planning assistant that can converse in spoken natural language with a person to create, discuss and evaluate various plans involving freight shipments by train. No prior training on how to interact with the system is required.

RIACS Seminar #126

Date: April 22, 2002

Title: "Model Checking, Program Analysis and Constraint Databases"

Speaker(s): Supratik Mukhopadhyay

Affiliation(s): University of Pennsylvania

Abstract:
Bugs in unverified (software) systems can cause disasters ranging from rebooting a PC to the failure of a space mission. Model checking is an automatic technique for verifying systems in which a desired behavior of a system is verified over a given system (the model) through exhaustive enumeration of all states reachable by the system and the behaviors that traverse through them. Program analysis refers to the technique(s) of automatically ascertaining information about a program without actually running the program. Constraint databases tightly integrate database and constraint solving methods therby bridging the gap between efficient, declarative database programming and efficient constraint solving.

We establish connections between the seemingly different fields of model checking for infinite state systems, program analysis and constraint databases. This connection allows us to derive uniformly solution large number of problems in verification of software. In particular, we derive uniformly, symbolic and (in most cases) local algorithms for interprocedural dataflow analysis, points-to analysis, aliasing analysis, automatic checking of array bound violation and other memory errors in C programs as well as automatic verification of safety and liveness properties of embedded software. I will also show how to seamlessly integrate deductive reasoning and abstract interpretation techniques within our methodology. The combined "framework" is used to derive "lightweight" tools for automatically verifying (rather falsifying = finding bugs in) software. I will share my experiences in developing and using a tool based on the described methodology for automatically discovering bugs in C programs.

Speaker's Bio:
Supratik Mukhopadhyay received his PhD in Computer Science from the Max Planck Institute for Computer Science, Saarbruecken, Germany, in May 2001. He started as a postdoctoral researcher at the Department of Computer and Information Sciences at the University of Pennsylvania from June 2001. His research interests include program analysis, embedded systems, software engineering and distributed systems.

RIACS Seminar #125

Date: April 18, 2002

Title: "Long Day's Drive: Advanced IT Enabled Long Distance Rover Traverse of Mars"

Speaker(s): Michael Sims

Affiliation(s): NASA Ames Research Center, Code IC/Center for Mars Exploration

Abstract:
NASA is expected to soon call for proposals for Mars missions for launched in 2007. A team including Ames, CMU and Ball Aerospace will propose a long distance rover to explore the the polar layered deposits in the northern region of Mars. In this talk I will describe that proposed mission, called Long Day's Drive, and will describe how clever design decisions allow mobility of 100 km or more. Central to this approach is relatively modest capabilities for rover 'self safing' autonomy. We need to have great confidence in the reliability of these self safing systems. Additionally, for Long Day's Drive some autonomy elements are more important and some less important than for the currently planned Mars Smart Rover which is to launch in 2009. Due to Long Day's Drive's long traverse over the polar layered deposits it will be possible to gather information that is otherwise obtainable by very deep drilling. We gather that information with considerably less complexity than drilling systems require. In this talk I will discuss the mission simplifications that have led to more autonomy with less effort and I will discuss those areas where advanced autonomy can significantly enhance the science return.

RIACS Seminar #124

Date: April 12, 2002

Title: "Testing Logistic Control Systems through Advanced Use of Simulation"

Speaker(s): Alexander Verbraeck

Affiliation(s): Delft University of Technology

Abstract:
Control systems for logistic and transport systems are extremely complex. Currently control systems are usually fully tested for the first time at the shop floor after commissioning. This usually results in a lot of costly failures that occur at the start-up stages of control systems, after all investments in the physical equipment have taken place. In our research, simulation plays an extended role that simulation can play in testing of fully automated logistic systems and their control systems before commissioning. We follow a three-step approach in testing both logistic and control systems, where a component based model of the control system and physical system plays an important role. A simulated control system is used to control simulated, emulated, and real prototypes of logistic resources. Until now, we tested our approach with different simulation COTS packages such as Simple++ / eM-Plant, AutoMod, and Arena. For one project for real-time control of free-ranging vehicles the control system has been implemented in all three simulation packages to control logistic resources at a vehicle TestSite. The TestSite is a special laboratory for testing new technologies in logistic automation, where we have 10 scale model vehicles of 2 metres long, 3 full truck size vehicles of 5-6 metres long, and loading and unloading docks. The simulation based control system enabled us to test the control strategies under different circumstances.

Speaker's Bio:
Alexander Verbraeck has an MSc in mathematics (cum laude, 1987) and a PhD in computer science (1991) from Delft University of Technology in the Netherlands. He worked as assistant professor in information systems until 1995, when he was appointed associate professor in the systems engineering group of the faculty of Technology, Policy and Management (TPM) of TU Delft. Current research focuses on complex distributed systems such as supply chains, real-time control and emulation of equipment using simulation, and on the development of generic, object oriented simulation libraries. Alexander has been the chair of the European Board of the Society for Computer Simulation International for five years and he is a member of ACM, IEEE-CS, INFORMS, AIS, and EuroSim. His research results have been published and presented at many international conferences. In 2001, Alexander Verbraeck was the department chair of the Information, Communication and Systems department and he served on the Faculty Board. From January 2002, Alexander also has an appointment as part-time research professor in the R.H. Smith School of Business of the University of Maryland in the Logistics, Business, and Public Policy Department.

Faculty of Technology, Policy and Management Systems Engineering Group
P.O. Box 5015, 2600GA Delft, THE NETHERLANDS
a.verbraeck@tbm.tudelft.nl
http://www.tbm.tudelft.nl/webstaf/alexandv

RIACS Seminar #123

Date: April 11, 2002

Title: "Formal Methods at IRST: Symbolic Model Checking and its Applications"

Speaker(s): Alessandro Cimatti

Affiliation(s): IRST (Institute for Scientific and Technological Research) in Trento, Italy

Abstract:
Goal of this talk is to overview the research lines and the technology transfer projects of the Formal Methods group at IRST, Trento, Italy. These activities are based on the use of symbolic model checking, a formal verification technique that allows for the analysis of digital systems represented as Finite State Machines.

In the talk, I will first describe NuSMV, a software platform for symbolic model checking developed at IRST as an OpenSource project. NuSMV combines the traditional use of Binary Decision Diagrams and the recently introduced Bounded Model Checking technology, based on propositional satisfiability. Then, I will show how symbolic model checking techniques are being extended and applied to the problems of Planning in Nondeterministic Domains, Safety Analysis, and Multi-agent reasoning.

Speaker's Bio:
Alessandro Cimatti is a senior research at IRST (Institute for Scientific and Technological Research) in Trento, Italy, where he is the head of the Formal Methods group. He participated in and led several industrial technology transfer projects aiming at the effective integration of Formal Methdos within the development process of safety-critical systems and commercial controllers. His research interests include Symbolic Model Checking and its applications to Planning in Nondeterministic Domains and to Multi-Agent reasoning.

RIACS Seminar #122

Date: April 4, 2002

Title: "Artificial Intelligence Techniques for Large-Scale Surveys of Space Science Data"

Speaker(s): Paul Gazis

Affiliation(s): San Jose State University Foundation

Abstract:
Many problems in space physics require large-scale surveys of extensive data sets to identify and classify qualitative features such as shocks, discontinuities, energetic particle enhancements, or specific types of spectra. Such surveys can be difficult to accomplish using conventional programming techniques and the manpower requirements associated with direct physical examination of the relevant data sets can be prohibitive. Artificial Intelligence (AI) techniques represent a potential solution to this problem. We have systematically applied and evaluated mature AI approaches, such as expert systems and several representative neural network architectures, to evaluate their suitability to perform large-scale surveys of solar wind plasma and interplanetary magnetic field (IMF) data. Several techniques were found to be particularly promising. Our results suggest that it should be possible to use AI techniques to identify and classify a broad range of different types of phenomena in large spectral data sets and extended time series that would be difficult or impossible to examine using any other means.

Speaker's Bio:
Dr. Gazis is a research associate at the San Jose State University Foundation. His thesis, from the Massachusetts Institute of Technology, was on solar wind evolution. Dr. Gazis worked at MIT Lincoln Laboratory to apply AI techniques to analyze images from infra-red laser radars and at the NASA Ames Space Science Division using conventional techniques to analyze solar wind data from the Pioneer and Voyager spacecraft. His interests involve the application of AI and statistical methods to the analysis of large space science data sets.

RIACS Seminar #121

Date: April 3, 2002

Title: "Neural Networks for Modeling and Control of Dynamic Systems"

Speaker(s): Dr. Atul G. Kelkar

Affiliation(s): Iowa State University

Abstract:
This talk is aimed at presenting some analytical, simulation, and experimental results on control of linear and nonlinear dynamic systems using neural network-based model predictive control. In particular, a class of model predictive controllers, namely, Generalized Predictive Controller (GPC) is used in this work. A brief background of neural networks with particular emphasis on their generalization capability that is central to their use in modeling and control of dynamic systems is given. Results related to the approximation accuracy of neural network and corresponding error bounds are discussed. Basic formulation of GPC control law for linear as well as nonlinear system is given. For linear unconstrained systems, conditions on GPC control parameters for ensuring closed-loop stability are given. A finite horizon stable GPC control strategy based on infinite horizon LQ performance function is proposed. For linear plants with uncertainties robustness conditions are derived for GPC control law. For control of uncertain linear systems and nonlinear systems, a neural network-based GPC architecture is presented. For uncertain linear systems, neural network is used in on-line learning configuration to account for plant uncertainties. For nonlinear systems affine in control, the nonlinear prediction equations are obtained for neural network-based predictor and closed-form of GPC control law is derived. The conditions for weight update that will ensure convergence of weights to optimal weights are given. Several experimental results are given to demonstrate the effectiveness of neural network-based GPC control schemes. Finally, open research issues in the area of NGPC are discussed and future research plan is presented.

RIACS Seminar #120

Date: March 28, 2002

Title: "Towards Component Retrieval in Constructive Logic"

Speaker(s): Ewen Denney

Affiliation(s): University of Edinburgh

Abstract:
The gazing technique was proposed in 1992 as a solution to the problem of lemma use in automated theorem proving. The idea is to consider a hierarchy of abstraction spaces, prove a goal by rewriting in an abstract space, then gradually refine this plan in concrete spaces, patching where necessary. In this talk, I discuss work in progress which aims to extend and apply this technique, in a constructive higher-order setting, to the problem of component retrieval. This suggests an interesting framework for combining deductive-based retrieval and adaptation. This is being carried out in the context of a project which formalises part of the Java Card virtual machine, and we will describe some aspects of this.

Speaker's Bio:
Dr Ewen Denney is a research fellow in the Division of Informatics at the University of Edinburgh. His thesis, from Edinburgh's Laboratory for the Foundations of Computing Science, was on a theoretical framework for formal program development. He has since worked on formal methods and theorem proving projects, in France and Hong Kong. His interests are in semantics, program synthesis, theorem proving, and their tool support and mathematical foundations.

RIACS Seminar #119

Date: March 21, 2002

Title: "Earth Science Technology Office (ESTO) Status and Plans"

Speaker(s): Dr. Walt Brooks

Affiliation(s): RIACS, NASA Ames Research Center

Abstract:
ESTO manages the development of advanced technologies and applications that are needed for cost-effective Earth Science missions. ESTO plays a major role in shaping ESE research and application programs of the future. The office is organized into three major areas including Information Systems. The objectives of the Advanced Information Systems Program area is to identify, develop and (where appropriate) demonstrate advanced information system technologies. These deemonstrations should: reduce the risk, cost, size, and development time of Earth Science Enterprise (ESE) space-based and ground-based information systems, increase the accessibility and utility of Earth science data, and enable new Earth observation measurements and information products.

This talk will address the overall approach, objectives, upcoming initiatives and solicitations relevant to Ames Information System Research and summarize the results of a recent workshop designed to identify technology drivers.

Speaker's Bio:
Dr. Walt Brooks is assigned to the level II NASA HQ Earth Science Technology Office at GSFC. His work focuses on Information Technology in support of future Earth Science Enterprise Missions. Previous work has included management of the Ames Supercomputer Facilities (NAS) and Program /Project management of several IR Space Astronomy missions.

RIACS Seminar #118

Date: March 21, 2002

Title: "The Role of Simplicity in Learning Theory"

Speaker(s): Dr. David McAllester

Affiliation(s): AT&T Laboratories-Research

Abstract:
Science values simplicity. All other things being equal, a simple explanation is preferable to a complex one. Bayesians assign higher prior probability to simple theories. But is it really true that a simple theory is a-priori more likely than a complex one? It turns out that one can justify a preference for simplicity independent of Bayesian assumptions. The justification involves only the law of large number and the observation that the number of simple theories is limited. This talk will present this justification and go on to describe more general "laws of large numbers" that justify more sophisticated methods of evaluating the accuracy of predictive rules. Applications to statistical natural language modeling will be discussed.

Bio Sketch:
David McAllester received his Ph.D. from the Massachusetts Institute of Technology AI Laboratory in 1987. He was on the faculty of the Cornell Computer Science Department for the 87/88 academic year and served on the faculty of the MIT AI Laboratory from 1988 to 1995. Since 1995 he has been a principal member of the technical staff at AT&T Laboratories-Research. David McAllester's research interests include knowledge representation, automated reasoning (automated theorem proving and formal verification), static analysis of computer programs, computer game playing (computer chess algorithms), constraint satisfaction algorithms, Bayesian networks, reinforcement learning, general PAC bounds in the theory of machine learning, and natural language modeling.

RIACS Seminar #117

Date: March 7, 2002

Title: "Collaborative Systems for the Distributed Management of Knowledge in Engineering Contexts"

Speaker(s): Dr. David Bell

Affiliation(s): Xerox PARC

Abstract:
There are few organizing structures more ubiquitous in modern organizations than standardized procedures and processes. For example, in field service engineering contexts, standard procedures are used in an attempt to optimize machine diagnosis and repair with respect to factors such as parts and labor costs; and in research & development (R&D) contexts, standard processes are used in an attempt to optimize coordinated work with respect to factors such as technology and market risk, product quality and cost, and program schedule.

The Internet is enabling new forms of collaborative systems that address these same organizational goals, but do so through more distributed means. Eureka and Sparrow Web (TM) are two collaborative systems developed at Xerox PARC that take advantage of the social nature of the Internet to support distributed management of knowledge in engineering contexts. Eureka has been used to support the work practices of field service technicians, and Sparrow Web(TM) has been used to support the work practices of R&D scientist and engineers. Both systems were developed using a human-centered and participatory design methodology that involved iteration between understanding actual work practices and co-designing and implementing collaborative systems.

This talk will contrast the traditionally centralized approaches for managing knowledge using standardized procedures and processes, with the more distributed approaches enabled through Eureka and Sparrow Web.

Speaker's Bio:
David Bell has been a member of the research staff at Xerox, where he was a core member of the research teams that invented the initial Eureka system and the latest version of the Sparrow Web system. During his ten years at PARC he worked in the Scientific & Engineering Reasoning Area of the Systems & Practices Laboratory, and conducted research in the areas of expert and case-based systems, knowledge and information management, and participatory human-centered design. David recently led a national research program on enterprise learning and knowledge sharing in the National Science Foundation sponsored Center for Innovation in Product Development at MIT. David received his Ph.D. from Cornell University, with a dissertation on "Product Development Process Dynamics."

RIACS Seminar #116

Date: March 7, 2002

Title: "Brahms: An Overview of the Tool and Current Research"

Speaker(s): Maarten Sierhuis

Affiliation(s): RIACS, NASA Ames Research Center

Abstract:
In this talk I present an overview of the current state of the Brahms multiagent modeling and simulation environment, and will discuss how we are using Brahms in our NASA research. Since my last Brahms presentation at ARC, the Brahms virtual machine has been fully integrated with the Java language, which now allows us to use Brahms not only as a simulation environment, but also as an intelligent agent development environment. I will discuss these new capabilities and will also discuss our progress with integrating Brahms with the KAoS agent framework from IHMC/UWF and Boeing Corp., allowing Brahms to run agents distributed in a networked environment. I will also present the first release of our new Brahms Interactive Development Environment, as well as our Brahms Virtual Reality Environment effort with DigitalSpace, Inc., integrating Brahms with DigitalSpace's OWorld and web-based virtual 3D worlds using AdobeR AtmosphereT.

Speaker's Bio:
Maarten is a Senior Research Scientist at RIACS/USRA, NASA Ames Research Center, Moffett Field, CA. He leads the Brahms Project in the Work Systems Design & Evaluation group (headed by Dr. William J. Clancey). Before coming to RIACS in 1998, Maarten was a member of technical staff, first in the Expert Systems Laboratory and later in the Work Systems Design group, at NYNEX Science & Technology. Maarten has a Ph.D. in Social Science and Informatics from the University of Amsterdam in The Netherlands, and an Engineering degree in Informatics from the Polytechnic University of The Hague, The Netherlands. His research interests lie in understanding the essence of human-centered computing, a new and upcoming multi-disciplinary field that could provide a leap in how information systems are designed and implemented. His view on this new field is influenced by his current research on modeling and simulating work practices in human organizations, as well as his past industry experience in developing knowledge-based systems attempting to make human organizations more effective and efficient.

RIACS Seminar #115

Date: February 21, 2002

Title: "Dialogue Design Strategies in Spoken Conversational Systems"

Speaker(s): Stephanie Seneff

Affiliation(s): Spoken Language Systems Group at the Laboratory for Computer Science, MIT

Abstract:
The Spoken Language Systems group at MIT's Laboratory for Computer Science has been developing mixed-initiative dialogue systems for spoken access to information for over a decade. This talk will focus on two of our most recently developed systems, the Mercury flight reservation system and the Orion system for task delegation. The latter system is able to call a user back at a prescribed time with pertinent information such as a traffic report or a flight arrival status. The emphasis of the talk will be on aspects concerned with dialogue design, including control, user feedback, confirmation, and error recovery. The overall system framework, which makes use of the Galaxy Communicator architecture, will also be discussed. We hope to be able to give a live demonstration of one or both of the systems.

Date: February 21, 2002

Title: "Enabling the Semantic Web: The Role of Metadata, Semantics and Domain Specific Ontologies"

Speaker(s): Dr. Vipul Kashyap

Affiliation(s): Technical Director of the ELBook project, Stanford Medical Informatics

Abstract:
We propose a vision of the Semantic Web, where domain specific ontologies and vocabularies will be the core infrastructure, as well as the basis of new metaphor for people and applications to interact with the web. We first illustrate with examples of how ontological concepts can be mapped to heterogeneous data, such as structured databases and textual databases. We then discuss the critical problem of inter-ontology interoperation, the core technology required to make the vision above a reality. Interoperation across ontologies requires queries to be re-written using inter-ontology relationships leading to translations that may not be semantics preserving. We present a novel approach for estimating loss of information due to the change in semantics. Measures for loss of information are defined based on intensional information; as well as on well established metrics like "precision" and "recall" based on extensional information. These are then used to select results from multiple translations across multiple ontologies. In this talk, we establish the critical role of metadata and domain specific ontologies in developing the semantic web infrastructure.

Speaker's Bio:
Vipul Kashyap works in the areas of Information and Knowledge Integration and Management, E-commerce and Semantic Web technologies. He has been active in the Semantic Web research community, and has organized panels and workshops on related topics. He has worked at R&D Labs of MCC and Telcordia Technologies (formerly known as Bellcore) on issues related to Information Integration and Agent based infrastructures and is the recipient of a Ph.D. from Rutgers University. His research interests are: interoperation across multiple domain ontologies on the Semantic Web and issues of loss of information as a consequence of ontology mismatches. Currently, he is investigating the feasibility of using sociological approaches for creating and evolving knowledge on the Semantic Web. Vipul has recently published a book on Information Brokering, has participated in panels, has been a member of conference program committees. He has published around 40 research articles and papers at various conferences and prestigious journals.

Date: February 20, 2002

Title: "Semi-Automated Indexing Using Domain-specific Semantics for High Precision Information Retrieval"

Speaker(s): Dr. Dan Berrios

Affiliation(s): Technical Director of the ELBook project, Stanford Medical Informatics

Abstract:
The use of full text resources like textbooks is frequently neither straightforward nor expedient. Faced with an urgent information need, a reader often must rely on manual inspection of a table of contents or alphabetized keyword index to guide her search. Indexes that would allow scientists to retrieve the information they need from text sources more rapidly and with greater precision must contain more knowledge than merely the location of the beginning of textbook sections or the numbers of pages on which one or two concepts are discussed. Entries in these indexes must mirror the questions that drive readers to use the text source to seek knowledge. Furthermore, these indexes must point the reader to more specific locations in the text.

We have developed a system, ISAID (Internet-based Semi-automated Indexing of Documents), to generate electronic indexes for HTML documents that are more detailed and more useful to readers. ISAID is part of ELBook, an integrated system for high-precision information retrieval from full-text resources. Users of ISAID see indexes proposed by the system, based on natural language processing of documents using domain-specific semantics from the Unified Medical Language System. In this seminar, I will discuss the design and implementation of ISAID and ELBook, including a controlled evaluation of ISAID's methods, in which users were timed and the indexes they generated compared.

Speaker's Bio:
Dan Berrios is Technical Director of the ELBook project at Stanford Medical Informatics. His research interests include collaborative information management, digital libraries and publishing, and natural language processing and web ontologies for information indexing and retrieval. He received a Bachelors degree in mathematics and biochemistry from Brown University in 1985, an M.D. from the University of California, San Francisco and Masters of Public Health in epidemiology and biostatistics from the University of California, Berkeley in 1990, and a Ph.D. in biomedical informatics from Stanford University in 2001.

---

Date: February 8, 2002

Title: "AWARE: Interpreting and Presenting Weather Data for Aviation Decision Making"

Speaker(s): Dr. Serdar Uckun

Affiliation(s): Director of Advanced Technology, Blue Pumpkin Software

Abstract:
A significant percentage of general aviation accidents and fatalities are attributable to weather. Ironically, most weather-related aviation accidents are inherently preventable by simply not launching a flight into potentially hazardous conditions. Although a variety of tools are available for disseminating raw weather data to pilots, accurately interpreting weather data in the context of a mission profile remains an art mastered only after thousands of hours of flying experience. In 1998, NASA launched a program named AWIN (Aviation Weather INformation) to develop technologies that may help reduce weather-related aviation fatalities. One of the AWIN programs, AWARE is a four-year effort jointly funded by Rockwell and NASA LaRC to interpret and present aviation weather data in order to assist pilots with aviation decision making tasks. In this talk, I will discuss the progress made in the AWARE program and expand on some of the research issues common to data-rich, knowledge-poor decision support problems.

Speaker's Bio:
Serdar Uckun is the Director of Advanced Technology at Blue Pumpkin Software, an enterprise software company focusing on scheduling and optimization applications for workforce management. Prior to joining Blue Pumpkin in 2000, he was Assistant Director and Manager of the Intelligent Systems Department at Rockwell Science Center, Palo Alto, CA. He served as the Program Manager for AWARE between from its inception in 1998 until early 2000. He has an M.D. from Ege University, Izmir, Turkey, an M.S. in Biomedical Engineering from Bogazici University, Istanbul, Turkey, and a Ph.D. in Biomedical Engineering from Vanderbilt University, Nashville, TN. Between 1992 and 1994, he received postdoctoral training in Computer Science at the Knowledge Systems Laboratory at Stanford University. His research interests include decision making under uncertainty, situational awareness, and scheduling.

Date: February 7, 2002

Title: "MarsNet Middleware Services for Remote Exploration"

Speaker(s): Dr. Norman Lamarra

Affiliation(s): NASA Jet Propulsion Laboratory

Abstract:
Recently, JPL's Center for Space Mission Information and Software Systems sponsored a study focused on issues and architectures for future space-based networks, focusing primarily on MarsNet. It studied mission and science challenges, and sketched an approach to addressing some of these challenges via modern IT, (processors, operating systems, software architectures and middleware services). A major recommendation of that study was that middleware services be developed to support future interplanetary networks (particularly for in-situ spacecraft), and that these could be deployed incrementally over several missions to the mutual benefit of all such missions.

Early work has begun with a prototype messaging middleware constructed with three "layers": application, middleware, and space protocol. This is intended to demonstrate separability that relieves applications from dealing with vagaries of the space communication. The prototype middleware is derived from a messaging API developed by JPL for the U.S. Marines, in which dispersed assets are required to communicate information over heterogeneous (and intermittent) networks, including low-bandwidth radios.

We refer to all layers between application and data link layer as "middleware", and preach the "service" approach to developing, deploying, operating, and evolving such middleware. Such middleware is in wide commercial deployment; for example, CORBA provides a set of distributed object services (naming, security, etc.) built upon (pluggable) lower-layer transport services (e.g., TCP/IP). Java also provides similar capabilities; for example, JINI and JXTA. With this background, this "messaging" prototype represents the necessary first step toward building a set of "middleware services" that are space-deployable and are designed ab initio to be combined into progressively higher-level capabilities. The next most important step proposed is a data-management prototype, leveraging Object-Oriented Data Technology infrastructure developed by JPL's Enterprise Data Management team. It is envisioned that evolution of such services would occur gradually over several years (and multiple missions), but we believe that resultant ubiquitous and rich interplanetary network infrastructure could be enabling for future space exploration.

Speaker's Bio:
Norm Lamarra is a Principal software systems engineer in JPL's Engineering & Communications Infrastructure section. He received the Ph.D. degree from UCLA in System Science in 1982. Prior degrees earned were M.Sc. in Radar Technology (1974) and B.Sc. in Mathematical Physics (1973), both from the University of Birmingham, UK. During the 25 years he has been in the U.S., Dr. Lamarra has worked in the fields of radar system analysis, real-time signal processing, adaptive antenna arrays, simulation and modeling for engineering and physiological systems, and software systems. He joined JPL in 1994.

Date: February 4, 2002

Title: "ScholOnto: Towards a Tool for Distributed Scientific Discourse"

Speaker(s): Simon Buckingham Shum

Affiliation(s): Knowledge Media Institute, Open University, UK

Abstract:
Digital libraries (DLs) are gradually becoming standard resources for researchers, but while useful, these also flood us with even more information than we already have to deal with. There remains a yawning gap in the scientist's digital toolkit: tools to track ideas and results in a field, and tools to express and analyse one's understanding of their significance. This is not surprising in that we're talking about meaning rather than datasets or published information. What is the significance of this idea in relation to others? According to whom? How does the expert community perceive this theory, model, language, empirical result? Where did this idea come from? What kind of evidence supports it, and challenges it? Are there different camps on this issue?

Scientific research in contested/poorly understood domains thus requires 'sensemaking' tools. Freeform annotation and discussions are the dominant solutions at present, but these generally have little structure and low status in scientific publishing, and are consequently perceived as very informal media.

The Scholarly Ontologies (ScholOnto) project [http://kmi.open.ac.uk/projects/scholonto], funded by the UK EPSRC, is exploring an alternative scenario. We are developing an ontology-based 'Claims Server' to support scholarly interpretation and discourse, investigating the practicality of publishing not only documents, but associated conceptual structures in a collective knowledge base. The system enables researchers to make claims: to describe and debate, in a network-centric way, their view of a document's contributions and relationship to the literature. It thus provides an interpretational layer above raw DLs (books; papers; datasets; software tools...). This contrasts with most DL/semantic web applications that require consensus on the structure of a domain, and an agreed metadata scheme that tries to iron out inconsistency, ambiguity and incompleteness. ScholOnto is all about supporting principled disagreement, conflicting perspectives, and the resulting ambiguities and inconsistencies, because they are the very stuff of research, and the objects of explicit inquiry.

In this presentation I'll describe where we've got to (1 year into a 3 year project), and demo the current version. I'll also illustrate how we're using another sensemaking tool, Compendium, to support ScholOnto's design (Compendium is also in use at ARC, by Maarten Sierhuis). Your insights and suggestions are very welcome.

Speaker's Bio:
Simon Buckingham Shum is a Senior Lecturer at the Open University's Knowledge Media Institute, a 70 strong R&D lab focused on the interaction between technology and knowledge. His background is in cognitive psychology, ergonomics and human-computer interaction. His central interest is in the human dimensions to knowledge media, with specific application to scholarly publishing and knowledge management. http://kmi.open.ac.uk/sbs

Date: January 15, 2002

Title: "Reading in the Digital Library: A Tale of e-Books in Action"

Speaker(s): Cathy Marshall

Affiliation(s): Microsoft Corporation

Abstract:
How will we read digital library materials? The answer to this question is far from straightforward; rather it has proven to be a matter that has polarized today's readers, writers, publishers, and librarians. Will we print digital documents as we need them, or will we be lured to new computer-based reading technologies like e-books? And if we are to read on a computer, what then? What might make it worthwhile to move away from a medium as convenient and malleable as paper? In this talk, I will take a work practice and technology in action-based look at reading, annotating, and collaborating over digital library materials to answer these questions.

Speaker's Bio:
Cathy Marshall is an Architect at Microsoft Corporation and an active participant in the international Hypertext, Digital Library, and WWW research communities. Her research lies in the disciplinary interstices of computer science, social science, and the arts. She was a long-time member of the research staff at Xerox PARC and an affiliate of the Center for the Study of Digital Libraries at Texas A&M University. Since 1983, Cathy has been working on computer support for intellectual work from the multiple perspectives of designer, theorist, (feral) self-styled ethnographer, and writer. Her current work investigates the use of mobile pen-based computers to support reading, research, collaboration, and annotation. Her homepage is http://www.csdl.tamu.edu/~marshall

Date: January 10, 2002

Title: "Genetic Algorithms and Schedule Optimization"

Speaker(s): Dr. Gil Syswerdal

Affiliation(s): i2 Corporation

Abstract:
In this talk, I will introduce genetic algorithms, and provide some computer demonstrations to show how they work on simple function optimization problems. I will then show how this technique can be applied to combinatorial optimization problems and to scheduling problems, which can be treated as constrained combinatorial optimization problems. Lastly, I will discuss what is required to take an optimization algorithm as the one described and create a successful commercial scheduling product.

Speaker's Bio:
Gil Syswerda received his Bachelors and Masters degrees in Computer Science from the University of Michigan in 1987, where he studied artificial intelligence in general and genetic algorithms in particular under John Holland, their inventor. In 1988, he joined the intelligent systems group at BBN Systems and Technologies, where he worked on scheduling and machine vision systems. In 1993, he co-founded Optimax Systems to produce commercial production scheduling systems. Optimax was purchased by i2 Technologies in 1997. Mr. Syswerda is currently a strategic technology and business advisor to i2.

Date: January 9, 2002

Title: "Adaptive and Parallel Discontinuous Galerkin Methods for Hyperbolic Systems"

Speaker(s): Joseph E. Flaherty

Affiliation(s): Scientific Computation Research Center, Rensselaer Polytechnic Institute

Abstract:
The discontinuous Galerkin method (DGM) provides an appealing approach to address problems having discontinuities, such as those that arise in hyperbolic conservation laws. Originally developed for neutron transport problems, the DGM has been used to solve both ordinary and partial differential equations. The DGM may be regarded as a way of extending finite volume methods to arbitrarily high orders of accuracy. The solution space is a piecewise continuous (polynomial) function relative to a structured or unstructured mesh. As such, it can sharply capture solution discontinuities relative to the computational mesh. It maintains local conservation on an elemental basis. Regardless of order, the DGM has a simple communication pattern to elements with a common face that makes it useful for parallel computation. It can handle problems in complex geometries to high order. And, it is useful with adaptivity since interelement continuity is neither required for h-refinement (mesh refinement and coarsening) nor p-refinement (method order variation).

We describe several aspect of the method including basis construction, data structures, flux evaluation, solution limiting, local time stepping, and a posteriori error estimation. We further describe a framework for controlling parallel adaptive computation. The parallel data management system can handle high-order techniques and maintain a dynamic load balance in homogeneous and heterogeneous computing environments. Results of serial and parallel computations are are presented for unsteady compressible flow problems involving instabilities and other complex two- and three-dimensional phenomena.

Speaker's Bio:
Joseph E. Flaherty is an Amos Eaton Professor in the Department of Computer Science and Mathematical Sciences at Rensselaer Polytechnic Institute of Troy, NY. He studied Aeronautical Engineering and received his Ph.D. in applied mechanics at the Polytechnic Institute of Brooklyn. He is currently Dean of the school of science at Rensselaer. Dr. Flaherty has researched and published extensively in the areas of computational science and numerical analysis such as adaptive finite analysis, adaptive local refinement, adaptive discontinuous Galerkin technique, grid generation and adaptive algorithms, error estimation and diffusion modeling. He also developed the degree programs at Rensselaer in Computational Science and Engineering and in Numerical Analysis.

Contact: flaherje@cs.rpi.edu

Date: December 13, 2001

Title: "Information Mediation: Integrating Information from Multiple Online Information Sources"

Speaker(s): Dr. Naveen Ashish

Affiliation(s): IBM Silicon Valley Laboratory

Abstract:
NOTE: This is an encore presentation, previously given on Oct. 23, 2001. Due to terrorist attacks, the former presentation was held off base, therefore this is an opportunity for those who missed that one to be able to attend.

This talk centers around "Information Mediators", which are software systems that provided integrated query access to multiple distributed information sources such as databases or Web sources. I will begin with an introduction to such systems, and provide an overview of several interesting research issues in building such systems. I will then discuss in detail an approach to optimizing the performance of these systems by locally storing data at the mediator side. Optimizing performance is the problem that I had addressed for my doctoral dissertation. I will conclude with talking about various applications of information mediators, future work in building mediators systems and finally, an overview of commercial ventures in this area.

Speaker's Bio:
Naveen Ashish holds a PhD in Computer Science from University of Southern California, Los Angeles. He is currently pursuing research in Life Sciences Development at the IBM Silicon Valley Laboratory, San Jose. He has been an assistant Professor in the Department of Computer Science at University of Georgia, Athens. His interests lie mainly in the area of information management, particularly information agents and information gathering and integration on the World Wide Web. His research background includes work and contributions in the areas of information mediators, integration of data from heterogeneous data sources and extraction and integration of data from semi-structured Web sources. He has worked in areas such as query planning and optimization for information mediators, information modeling and also semi-automatic wrapper generation for extracting data from Web sources.

Note: Please see the homepage of the Information Agents Research Group, for an overview.

Contact: naveenashish@hotmail.com

Date: December 3, 2001

Title: "Using Knowledge to Save Work"

Speaker(s): Dr. Bob Young

Affiliation(s): SciComp, Inc.

Abstract:
Knowledge-based systems can save work both for end users, such as numerical modelers, and for the developers of the systems that help the users. For example, user interface components can be generated from the same knowledge bases that drive the performance of a system. One instance of how this can work comes from the domain of numerical simululation, which plays a very important role in many industries. SciComp was founded with the belief that we could help numerical modelers to produce high quality programs with much less effort on their part. SciComp's SciFinance product is a knowledge-based system that transforms concise problem descriptions, given in financial and mathematical terms, into running C-code simulations. The system has knowledge of programming, mathematics and finance. We took advantage of this environment to generate a number of different kinds of hyper-linked documents that are important components of the users' interface to SciFinance. They include reference documents, "design history" documents and viewers that offer different views of the same data, aimed at different kinds of users: mathematically knowledgable user vs system developer. Most of the documents were automatically generated, and hence a major labor saving, only possible beause of the knowledge-based environment.

Speaker's Bio:
My main interest is Human Computer Interaction: using information technology to make complex tasks easier for the people who do them. In pursuit of this goal, I have worked on problems of interest to geologists, petrophysical engineers, mechanical designers, test equipment designers and most recently, modelers with a need for numerical simulation. These problems have required large amounts of diverse knowledge for their solution.

SciComp Inc: (Austin, TX) Director of Technology, A founding employee of SciComp

SciComp products generate simulation programs from a very high-level specification language. The programs solve partial differential equations and stochastic differential equations. The technology is applicable in many fields, but has been initially targeted at finance. Among a variety of activities, I designed and implemented the Electronic Information System (EIS) for our product system.

Schlumberger: Aided ATE designers in developing simple ways to organize the information about a large set of software tools to help new employees get up to speed quickly without extensive human mentoring. Led a group that developed prototypes of CAD engineering tools that significantly improved both the speed and quality of customers' mechanical design process by augmenting traditional geometric represen