学术讲座

当前位置: 首页>> 科学研究>> 学术讲座>> 正文

光华讲坛—【鲁棒与随机优化系列讲座(五)】Stochastic Gradient Estimation
发布时间: 2021-04-30

主题【鲁棒与随机优化系列讲座(五)】Stochastic Gradient Estimation

主讲人北京大学彭一杰助理教授

主持人工商管理学院徐亮教授

时间2021年5月6日(星期四)14:00-15:00

直播平台及会议ID腾讯会议 会议ID:413 420 714

主办单位:工商管理学院科研处

主讲人简介:

Dr. Yijie Peng is an Assistant Professor in Guanghua School of Management and Institute for Artificial Intelligence at Peking University (PKU). He received his Ph.D. from the Department of Management Science at Fudan University and his B.S. degree from the School of Mathematics at Wuhan University. Many of his publications appear in high-quality journals including Operations Research, INFORMS Journal on Computing, and IEEE Transactions on Automatic Control. He is awarded the 2019 Outstanding Simulation Publication Award of INFORMS simulation society. He serves as an Associate Editor for Asia-Pacific Journal of Operational Research and IEEE Control Systems Society Conference Editorial Board. His research interests include stochastic modeling and analysis, simulation optimization, machine learning, data analytics, and healthcare.

彭一杰博士现任北京大学光华管理学院和人工智能研究所助理教授,此前在武汉大学数学系和复旦大学管理科学系分别获得学士、博士学位。研究成果发表于Operations Research, INFORMS Journal on Computing, IEEE Transactions on Automatic Control等国际学术刊物。2019年荣获INFORMS仿真协会杰出论文奖。现担任Asia-Pacific Journal of Operational Research期刊、IEEE Control Systems Society会议编委会副主编。主要研究领域为随机建模与分析、仿真优化、机器学习、数据分析和医疗健康。

内容提要:

Abstract: Stochastic gradient estimation is an actively studied field in simulation, because it plays a central role in gradient-based optimization and sensitivity analysis. Infinitesimal perturbation analysis (IPA) and the likelihood ratio (LR) method are two classic unbiased derivative estimation techniques. Traditional applications are in discrete event dynamic systems and financial engineering and risk management. Recently, the stochastic gradient estimation techniques have attracted attention in machine learning and artificial intelligence. A key challenge in stochastic gradient estimation has been handling discontinuities in sample performance for structural parameters, which arise in wide variety of applications including financial engineering, production/inventory management, and training artificial neural networks. In this talk, the speaker will present a new approach called generalized likelihood ratio (GLR) method, capable of dealing with a large scope of discontinuities in a general framework. GLR is applied to train artificial neural networks in examples of image recognition and reinforcement learning. The new method offers unbiased estimates for distribution sensitivities, which can be used to construct confidence intervals and regions for quantiles without batching, calibrate stochastic models without analytical likelihood, and solve quantile-based optimization efficiently. In these examples, certain robustness improvement can be achieved by using GLR.

随机梯度估计在仿真中有着广泛研究,在基于梯度的优化和灵敏度分析中扮演着重要角色。本次讲座将介绍一种新的随机梯度估计方法,称为广义似然率方法,该方法能有效处理非连续系统,在训练人工神经网络、强化学习等领域中具有良好理论性质。