MURI Semiars

MURI Seminars

Seminars in March and April, 1997

Seminar Abstracts

Tuesday, April 22nd, 1997
4-5:30pm
306 Soda Hall
Reinforcement Learning, Dynamic Games, & Intelligent Control

Nital S. Patel

Systems Engineering & Integration Laboratory
Institute for Systems Research
University of Maryland at College Park
Abstract:

In this talk I will present a dynamic game framework as a potential tool for studying issues arising in intelligent control problems. Here one views the controller and nature (uncertainty) as playing against each other. These dynamic games also play a key role in robust nonlinear control theory. Although, both intelligent and robust control deal with problems arising in coping with uncertainty, robust control assumes that the nature and effects of this uncertainty are known a priori. In contrast, intelligent controllers have to deal with uncertainty that has not been anticipated, such as arising out of changes in the environment, or the system itself. This requires that the controller not only be able to generate and update internal world models, but also carry out iterative improvements in its performance. Furthermore, this has to be done in finite time and with finite resources. Although, a general theory that addresses all these issues has yet to be developed, the dynamic game framework provides a stepping stone in this direction. During this presentation, I will talk about the role of randomization, results known for application of reinforcement learning to these randomized dynamic game problems (off-line, as well as, on-line), some open problems and key issues that still need to be addressed.

 Tuesday, April 15, 1997
4-5:30pm
306 Soda Hall

Planning, Control and Learning with Hierarchical Behaviors

Ron Parr
Department of Computer Science
UC Berkeley

This talk presents a new approach to planning and control problems in the Markov decision process (MDP) formalism. In this method partially specified machines can be used to express abstract policies or strategies. This provides a framework in which knowledge can be transferred across problems and in which component solutions can be recombined to solve larger and more complicated problems. Component machines can also be used to constrain the set of solutions that is considered for new problems, in many cases reducing large problems to much smaller ones. Unlike many hierarchical approaches based on state aggregation, our approach retains the Markov property at higher levels of abstraction. Therefore, it can build on the well-understood mathematical properties of MDPs, and can be seen as providing a link between MDPs and ``behavior-based'' or ``teleo-reactive'' approaches to control.

 Tuesday, April 8st, 1997
4-5:30pm
306 Soda Hall

Organizing Data: Supervised Categorization/Classification in
non-Metric Spaces

Daphna Weinshall*
New-York University
& The Hebrew University of Jerusalem
Abstract

The problem of supervised categorization/classification occurs in many areas ranging from cognitive psychology to gene matching and computer vision: it is the problem of constructing a representation reflecting similarities and differences between apriori classified data, such as groups of images. Pattern recognition and artificial neural networks have for the most part focused on classification in Euclidean feature spaces. However, the advent of robust techniques in application areas such as computer vision revealed the importance of non-metric similarity computation; in fact, most matching methods used in practice in computer vision are not Euclidean, and rarely even metric. Cognitive psychologists have long known that human judgment of similarities is also non-metric.

In this talk I will discuss the computational implications of non-metric matching techniques on supervised classification. I will describe many problems, and address a few of them. In particular, I will focus on the absence of transitivity as a major problem for data reduction, and present a simple algorithm to achieve data reduction in non-metric data sets. Simulations show the superior performance of our method under various conditions. I will discuss in detail a novel curve matching algorithm, which uses the robust median distance and is therefore non-metric.

----
* joint work with David Jacobs, NECI

 Tuesday, April 1st, 1997
4-5:30pm
306 Soda Hall

Visually Locating and Monitoring Moving Targets in Cluttered
Environments

Steven M. LaValle
Computer Science Department
Stanford University

Abstract:

Visibility constraints lead to new challenges for computing motion strategies when considered in addition to geometric constraints that are typical in basic motion planning. The recent study of such problems is motivated by our experimentation with mobile robots that behave as "autonomous observers," which can be used in applications such as tracking people or other robots to supply information for distributed virtual environments, monitoring processes in an assembly workcell, automatically moving cameras to avoid obstructions in a medical surgery site, or searching for potentially-hostile targets for surveillance. This talk will focus on two visibility-based planning problems. The first problem involves coordinating the motions of one or more pursuers that have omni-directional vision sensors to eventually ``see'' an evader that is unpredictable, has unknown initial position, and is capable of moving arbitrarily fast. The second problem involves computing motion strategies that attempt to maintain visibility of a moving target that may or may not be predictable, while optimizing additional criteria such as the total distance traveled. Several computed results and mobile robot

experiments will be presented. WWW info
--------

http://robotics.stanford.edu/~lavalle

 Tuesday, March 18, 1997
306 Soda Hall

Querying Multiple Sources across the Internet

Luis Gravano
Computer Science Department
Stanford University

Faculty Canditate

gravano@cs.stanford.edu
http://www-db.stanford.edu/~gravano

Abstract:

Information sources are available everywhere, both within the internal networks of organizations and on the Internet. To evaluate a query over many such sources and give the users the illusion of a single, large source, we need to perform three main tasks. First, we need to choose the best sources to evaluate a query. Second, we have to submit the query to these sources. Third, we need to merge the query results from the sources. I will first describe GlOSS, a scalable system that chooses the best document sources for a query. The GlOSS information about each source is orders of magnitude smaller than the source contents. I will also give an overview of the design of STARTS, an emerging protocol for Internet retrieval and search involving over 11 companies and organizations. Finally, I will address the problem of efficiently retrieving ranked information from sources. In particular, I will discuss the problem of merging ranked query results from sources with structured data, and optimization algorithms for queries over repositories of complex multimedia objects.
Tuesday, March 4th, 1997
4-5:30pm
306 Soda Hall

Range Estimation by Optical Differentiation

Hany Farid
University of Pennsylvania
GRASP Lab

Abstract:

We describe a novel formulation of the range recovery problem based on computation of the differential variation in image intensities with respect to changes in camera position (or aperture size). This method uses a single stationary camera and a pair of calibrated optical masks to directly measure this differential quantity. The subsequent computation of the range image involves simple arithmetic combinations, and is suitable for real-time implementation. Both the theoretical and practical implications of this formulation are addressed.

(Joint work with Eero Simoncelli)

[Top|Home]


[MURI Home|Website Managers|September 1996]