國立東華大學應用數學系

專    題    演    講

主講人：顏佐榕

        中央研究院統計科學研究所

講  題： Parameter clustering: Encouraging similarities between estimates via Euclidean distance regularization

時  間：104 年 03 月 27 日 (星期五)  15:00-17:00

地  點：理工一館 A324會議室

摘   要

In statistical estimation, one important goal is to obtain a model that has better ability in prediction but fewer parameters for interpretation. Such parsimony requirement leads statisticians to develop various techniques for reducing the effective number of parameters in the model. In this paper we propose a penalized estimation method to fulfill this requirement. The method aims to reduce the effective number of parameters by estimating parameters with identical values. It imposes $l_{2}$-norm penalty functions on differences between pairs of the parameters. Under this setting, the method is able to shrink the differences to zero, yielding identical estimates for the parameters. To numerically carry out the method, we first formulate the problem as a constrained optimization problem, and then solve the constrained optimization problem by developing an iterative algorithm based on the alternating direction method of multipliers. Simulation studies show that the method can simultaneously identify the number of effective parameters and deliver collaborative estimates for these parameters. We discuss several applications and a proposal for carrying out this method via distributed optimization.

※※※   歡      迎      參      加   ※※※se1040327 附檔：