Y. L. Tong

Georgia Tech

**October 24, 1996**

Doug Down

Georgia Tech

**October 31, 1996**

Tom Kurtz

University of Wisconsin

**November 1, 1996**

Reuven Y. Rubinstein

Technion, Israel

**November 7, 1996**

Walter Philipp

University of Illinois Urbana-Champaign

**November 14, 1996**

Raid Amin

University of West Florida

**November 21, 1996**

Serguei Foss

Novosibirsk State University and Colorado State University

In the second part of the talk, we consider the following problem. Let $Y \equiv Y_0 \equiv \{ X_n, n \geq 0 \}$ be a sequence of random variables. For $k=1,2, \ldots $, put $Y_k = \{ X_{k+n}, n \geq 0\}$ and denote by $P_k$ the distribution of $Y_k$. When does there exist a probability measure $P$ such that $P_k \to P$ in the total variation norm?

**December 5, 1996**

Minping Qian

Peking University, Beijing, China

**January 10, 1997**

Andre Dabrowski

University of Ottawa

After a brief decription of patch-clamp experiments and their associated data, we will provide an overview of the major approaches to the statistical analysis of current traces. The renewal-theoretic approach of Dabrowski, McDonald and Rosler (1990) will be described in greater detail, and applied to the analysis of data arising from an experiment on stretch-sensitive ion channels.

**January 16, 1997**

Christian Houdre

Georgia Tech

**January 30, 1997**

Ming Liao

Auburn University

(1) a random "rotation" which tends to a limit as time goes to infinity;

(2) an asymptotically deterministic flow;

(3) another random "rotation".

Using this decomposition, we may describe the random "sinks" and "sources" of the stochastic flow explicitly. Examples of stochastic flows on spheres will be discussed.

**February 6, 1997**

Dana Randall

Georgia Tech

**February 13, 1997**

Steen Andersson

Indiana University

This paper combines conditional independence (CI) restrictions with group symmetry (GS) restrictions to obtain the group symmetry conditional independence (GS-CI) models. The group symmetry models and the conditional independence models are thus special cases of the GS-CI models. A complete solution to the likelihood inference for the GS-CI models is presented.

Special examples of GS models are Complete Symmetry, Compound Symmetry, Circular Symmetry, Complex Normal Distributions, Multivariate Complete Symmetry, Multivariate Compound Symmetry, and Multivariate Circular Symmetry. When some of these simple GS models are combined with some of the simple CI models, numerous well-behaved GS-CI models can be presented.

**February 27, 1997**

Dimitris Bertsimas

Sloan School, MIT

The first approach aims to explore the dynamic character of the problem by considering the fluid model of the queueing network. We propose an algorithm that solves the fluid control problem based on infinite linear programming. Our algorithm is based on nonlinear optimization ideas, and solves large scale problems (50 station problems with several hundred classes) very efficiently.

The second approach aims to shed light on the question of how stochasticity affects the character of optimal policies. We use singular perturbation techniques from the theory of partial differential equations to obtain a series of optimization problems, the first of which is the fluid optimal control problem mentioned in the previous paragraph. The second order problem provides a correction to the optimal fluid solution. This second order problem has strong ties with the optimal control of Brownian multiclass stochastic networks. We solve the problem explicitly in many examples and we see that the singular perturbation approach leads to insightful new qualitative behavior. In particular, we obtain explicit results on how variability in the system affects the character of the optimal policy.

**March 5, 1997**

Paul Glasserman

Columbia Business School

The first part of this talk reviews positive results that support this strategy and examples of its potential for dramatic variance reduction. The second part shows, however, that the same approach can be disastrous even in very simple examples. For each negative example, we propose a simple modification that produces an asymptotically optimal estimator.

**March 13, 1997**

Hayriye Ayhan

Georgia Tech

**April 17, 1997**

Andrew Nobel

University of North Carolina, Chapel Hill

The procedure divides the available data into two parts. The first is used to select an empirical cover of each model class. The second is used to select from each cover a candidate rule with the smallest number of misclassifications. The final estimate is chosen from the list of candidates in order to minimize the sum of class complexity and empirical probability of error.

A distinguishing feature of the approach is that the complexity of each model class is assessed empirically, based on the size of its empirical cover.

**April 24, 1997**

Robert Adler

University of North Carolina, Chapel Hill

and Technion -- Israel Institute of Technology

In this talk I shall try to provide an introduction to superprocesses for the uninitiated, describing their basic structure, some basic results, and some interestimg open questions.

The talk will be followed by a 15 minute movie for anyone who wishes to stay.

**May 1, 1997**

Alex Koldobsky

University of Texas at San Antonio

**May 8, 1997**

Bok Sik Yoon

Hong-Ik University, Seoul, Korea & Georgia Tech

**May 15, 1997**

Jan Rosinski

University of Tennessee, Knoxville

**May 22, 1997**

Robert Cooper

Florida Atlantic University

Recently, it has been found that polling models themselves enjoy an analogous decomposition with respect to the server switchover times (or, in the manufacturing context, setup times), but for apparently different reasons. Furthermore, it has recently been discovered that some polling models exhibit counterintuitive behavior: when switchover times increase, waiting times decrease; or, equivalently, in the parlance of manufacturing, WIP (work in process) can be decreased by artificially increasing the setup times.

In this talk we give an overview of polling and vacation models, including some historical context. Also, using decomposition we "explain" the counterintuitive behavior, and identify it as a hidden example of the well-known renewal (length-biasing) paradox.

The talk will emphasize conceptual arguments rather than mathematical detail, and should be of interest to a general audience.

Unless otherwise noted, the seminar meets Thursdays at 3 PM in Skiles, Room 140. For further information, contact Jim Dai (dai@isye.gatech.edu) or Richard Serfozo ( rserfozo@isye.gatech.edu).

**May 29, 1997**

Takis Konstantopoulos

University of Texas, Austin

**June 5, 1997**

Alan F. Karr

National Institute of Statistical Sciences

This talk will describe a large, cross-disciplinary, multi-organization study, now in its first year, meant to define, measure and visualize code decay, to identify its causes (both structural and organizational), to quantify effects, and to devise remedies. Emphasis will be on the code itself and its change history as statistical data, and on tools to describe and visualize changes.

*Last updated: May 31, 1997 by J. Hasenbein ( johnny@isye.gatech.edu)*