报告人：牟必强，中科院数学与系统科学研究院
报告题目：Regularization methods for system identification
报告时间：2021年6月24日周四9:00-11:00
报告地点：腾讯会议 ID: 163 407 106
报告摘要：The classical system identification identification (maximum likelihood/prediction error methods) are not as reliable as expected when the data is short and/or has low signal-to-noise ratio and also encounter the difficulty for choosing a proper model order. Kernel-based regularization methods developed in the last ten years is a new paradigm for system identification, which consists of two core issues: kernel design and hyper parameter estimation. Kernel design encodes prior knowledge of the dynamical system to be identified (exponential decay, smoothness, etc) into the parameterization of the kernel, which plays a similar role in the classical model structure selection. Hyperparameter estimation uses the data to estimate the hyper-parameter for parameterizing the kernel, which actually tunes a suitable model complexity in a continuous way and is similar to the classical model selection. Input design is an important issue for further improving the accuracy of kernel methods.
报告人简介：牟必强，中国科学院数学与系统科学研究院副研究员。于2008年从四川大学获得工学学士学位，于2013年从中国科学院数学与系统科学研究院获得理学博士学位。牟必强博士研究兴趣包括动态系统建模与辨识，机器学习核方法，动力电池老化机理分析等。在自动控制领域 Automatica，IEEE Transactions Automatic Control、SIAM Journal on Control and Optimization，European Journal of Control 等期刊上发表多篇论文。