时 间：2025年6月10日 11:00-12:00
地 点：普陀校区理科大楼A1314
报告人：杨翔宇 山东大学博士后
摘 要：
We propose a penalty-based method for solving stochastically constrained simulation optimization problems with differentiable structure. We introduce a novel class of penalty functions and reformulate the original constrained problem as a sequence of unconstrained subproblems, which are approximately solved using a multi-timescale stochastic gradient descent framework. The method combines features of both exterior and interior penalty approaches: it acts like an exterior method to guide an infeasible solution toward the feasible region and transitions to an interior method once feasibility is achieved. In contrast to existing methods, our approach operates fully online, eliminates the need for explicit feasibility checks, and requires no auxiliary optimization subroutines during the iterative process. Under suitable assumptions, we establish almost sure convergence of the algorithm and derive an explicit convergence rate for the final iterate via a bias–variance decomposition of the stochastic gradient estimators. Numerical experiments on both synthetic and real-world datasets are performed to illustrate the practical effectiveness and robustness of the proposed method.
报告人简介：
杨翔宇，现为山东大学特别资助类博士后。他于2022年6月在复旦大学管理科学系获得博士学位，随后加入山东大学管理学院工作至今，期间曾获山东大学优秀博士后一等奖（10人/年）。他的主要研究方向为基于随机仿真的优化算法及其应用，相关成果已发表于Automatica、European Journal of Operational Research、IEEE Transactions on Automatic Control等。他的研究受到国家自然科学基金青年项目（C类）、中国博士后科学基金面上项目、山东省自然科学基金青年项目（C类）等资助。