题目：《Sequential Numerical Integration and Stochastic Optimization with Statistical Designs》
报告人：钱智光博士，美国威斯康星大学统计系及工业工程系助理教授
时间：12月30日下午2点
地点：曹光彪东楼502
Numerical integration and stochastic optimization are at the heart of many estimation problems in statistics, data mining and machine learning. Examples include marginalization of nonlinear likelihood functions, imputation of missing data in regularized variable selection procedures, maximum likelihood estimation of hierarchical models and posterior mode calculation of Bayesian graphical models.
The first part of my talk is devoted to sequential numerical integration with statistical designs. I will discuss several new sampling designs for accurately estimating multi-dimensional integrals in a sequential fashion. These designs are combinational in nature and are constructed by exploiting nested relationships in stratified permutations, difference schemes, linear codes and other discrete structures.
The second part of my talk deals with enhancing stochastic programming methods with space-filling designs. A stochastic program is an optimization problem in which the objective function is a multi-dimensional integral. The popular Sample Average Approximation method solves a stochastic program by constructing a sampling based approximation to the objective function of the optimization problem and then finding the solution of the approximated problem. Independent and identically distributed sampling is the prevailing choice for generating such an approximation. I will present a general theory of sample average approximations with space-filling designs. In particular, it is shown that, in terms of large deviations, sample average approximations with space-filling designs are far superior to those with identically distributed sampling.
Theoretical properties and numerical illustrations of the developed methods will be presented. The talk is based on joint papers with Ben Haaland and Qi Tang at the University of Wisconsin-Madison, and Mingyao Ai at Beijing University.
报告人简介：
钱智光博士现为美国威斯康星大学麦迪逊分校统计系和工业工程系助理教授。他在2006年从乔治亚理工大学获博士学位，2003 年从密西根大学获硕士学位，2000年从华东师范大学获本科学位。他的研究领域包括机器学习，随机优化，试验设计和计算机试验。他是2008年IBM教授奖获得者之一。（邮箱：peterq@stat.wisc.edu）
报告人：钱智光博士，美国威斯康星大学统计系及工业工程系助理教授
时间：12月30日下午2点
地点：曹光彪东楼502
Numerical integration and stochastic optimization are at the heart of many estimation problems in statistics, data mining and machine learning. Examples include marginalization of nonlinear likelihood functions, imputation of missing data in regularized variable selection procedures, maximum likelihood estimation of hierarchical models and posterior mode calculation of Bayesian graphical models.
The first part of my talk is devoted to sequential numerical integration with statistical designs. I will discuss several new sampling designs for accurately estimating multi-dimensional integrals in a sequential fashion. These designs are combinational in nature and are constructed by exploiting nested relationships in stratified permutations, difference schemes, linear codes and other discrete structures.
The second part of my talk deals with enhancing stochastic programming methods with space-filling designs. A stochastic program is an optimization problem in which the objective function is a multi-dimensional integral. The popular Sample Average Approximation method solves a stochastic program by constructing a sampling based approximation to the objective function of the optimization problem and then finding the solution of the approximated problem. Independent and identically distributed sampling is the prevailing choice for generating such an approximation. I will present a general theory of sample average approximations with space-filling designs. In particular, it is shown that, in terms of large deviations, sample average approximations with space-filling designs are far superior to those with identically distributed sampling.
Theoretical properties and numerical illustrations of the developed methods will be presented. The talk is based on joint papers with Ben Haaland and Qi Tang at the University of Wisconsin-Madison, and Mingyao Ai at Beijing University.
报告人简介：
钱智光博士现为美国威斯康星大学麦迪逊分校统计系和工业工程系助理教授。他在2006年从乔治亚理工大学获博士学位，2003 年从密西根大学获硕士学位，2000年从华东师范大学获本科学位。他的研究领域包括机器学习，随机优化，试验设计和计算机试验。他是2008年IBM教授奖获得者之一。（邮箱：peterq@stat.wisc.edu）
