报告时间：2019年12月16日 上午10:30

报告地点：工西417

报告人及单位：李勇祥（上海交通大学）

报告摘要：Large-scale Gaussian process (GP) models are becoming increasingly important and popularly used in the context of geostatistics, machine learning, simulation optimization, etc. However, the standard methods of GP models, the maximum likelihood estimation and the best linear unbiased predictor, are designed to run in a single computer whose computational power is often limited even for a computer in a super-computing center. Therefore, approximate alternatives that can use the power of multiple computers are in an increasing demand, such as the composite likelihood methods. However, those alternative methods in the literature offer limited options for practitioners, because most methods care more about computational efficiency than the statistical efficiency. In fact, there is lack of methods in the literature that can provide accurate solutions to parameter estimation and model prediction of large-scale GP applications for practitioners who can use a super-computing center. Therefore, we develop an optimal composite likelihood method in this paper that tries to minimize information loss in parameter estimation and prediction of large-scale GP models. We prove that the proposed composite likelihood prediction, called the best linear unbiased block predictor, has the minimum prediction variance under some conditions. Numerical examples show that both the composite parameter estimation and prediction method we proposed exhibit more accurate performance than their traditional counterparts under various cases.

报告人简介：李勇祥博士现为上海交通大学工业工程与管理系助理教授，于2019年在香港城市大学数据学科学院获得博士学位。他主要从事数据科学在复杂系统的优化设计，质量控制和故障检测中的不确定性研究，主要研究方向包括计算机试验设计与分析，统计质量控制，统计信号处理等。他在统计与工业工程领域著名期刊Technometrics、IEEE Transactions on Signal Processing等期刊发表多篇论文。