乐于分享
好东西不私藏

4月17日“软件新技术讲坛”学术报告 - 陈程 副教授

4月17日“软件新技术讲坛”学术报告 - 陈程 副教授

Robust Multi-agent Multi-armed Bandits: From Corruption-Resilience to Byzantine-Resilience

时间:2026年4月17日(星期五)15:00

地点:苏州校区南雍楼西122室

陈程,副教授

华东师范大学 软件工程学院

摘  要

Cooperative multi-agent multi-armed bandits (CMA2B) investigate how multiple agents collaborate to minimize the regret of a multi-armed bandit problem. Although this setting has been extensively studied, most existing algorithms remain vulnerable to various forms of adversarial manipulation. In this talk, we first introduce a CMA2B framework that is robust to adversarial corruption, where an adversary can corrupt the reward observations of all agents under a limited corruption budget. We then consider a more realistic scenario in which the adversary can attack only a small subset of agents. We show that in this case, the impact of adversarial attacks can be almost completely eliminated, and that the framework is inherently robust in the Byzantine setting, where an unknown fraction of agents may arbitrarily select arms and spread incorrect information.

报告人简介

陈程,华东师范大学软件工程学院副教授。于上海交通大学计算机科学与技术专业本科直博,博士期间在美国加州大学伯克利分校数学系公派联培一年,之后前往新加坡南洋理工大学数学系从事博士后研究工作。主要研究方向包括在线机器学习、最优化理论、强化学习以及矩阵近似。在机器学习领域的顶级会议和期刊上发表学术论文十余篇,入选上海市高层次青年人才计划,并获得国家自然科学基金青年项目资助。目前担任机器学习旗舰期刊JMLR的编委会审稿人 (Editorial board reviewer),并多次担任NeurIPS、ICML、ICLR等机器学习顶级会议的审稿人。
欢迎参加