时间：2024年12月26日（周四），上午9:30-11:30

地点：计算机与人工智能学院报告厅

Talk title: Deep Reinforcement Learning for Complex Decision-Making

Deep reinforcement learning proves to be an effective optimization method for resolving numerous complex decision-making problems. In this seminar talk, I will begin by providing a brief overview of deep reinforcement learning and the underlying model, Markov Decision Processes (MDP). Subsequently, I will delve into the challenges that make MDP hard to solve, addressing key issues such as the extremely vast state/action space, the reward alignment problem, and the policy generalization challenge. Furthermore, I will showcase examples of employing deep reinforcement learning to tackle challenging problems in video games, multiagent systems, and simulated robot controls. Towards the conclusion of the talk, I will introduce the ongoing challenges we are currently addressing, with a specific emphasis on combinatorial decision-making on graphs and efficient optimization methods for large models.

演讲题目: 深度强化学习在复杂决策中的应用

深度强化学习被证明是一种有效的优化方法，可解决众多复杂的决策问题。在本次研讨会中，我将首先简要介绍深度强化学习及其基础模型——马尔可夫决策过程（MDP）。随后，我将深入探讨使得MDP求解中存在的关键性问题，例如极其庞大的状态/动作空间、奖励对齐问题以及策略泛化的挑战。此外，我将展示使用深度强化学习解决视频游戏、多智能体系统和模拟机器人控制中的复杂问题的实例。在演讲的最后，我将介绍我们目前正在解决的持续性挑战，特别是关于图上的组合决策问题和大型模型的高效优化方法。

Short bio:

Dr. Mingfei Sun is a professor in Machine Learning in the Department of Computer Science and leads the Agent Lab at the University of Manchester (UoM). His research focuses on deep reinforcement learning and deep learning for decision-making. Dr. Sun serves on the Admission Panel for the AI Centres for Doctoral Training Programme in Complex Decision-Making and holds the position of Student Experience Director. Prior to joining UoM, Dr. Sun worked as a researcher at Microsoft Research Cambridge, focusing on foundation models, and as a research associate at the University of Oxford, specializing in multi-agent systems. Dr. Sun's work has been recognized at leading machine learning and artificial intelligence conferences, including NeurIPS, ICLR, KDD, AAAI, IJCAI, and AAMAS. He has received prestigious honors such as the Best Paper Award at AAMAS and the Google Summer of Code Award.

简介: 

孙明飞博士在曼彻斯特大学 (University of Manchester, UoM) 计算机科学系从事机器学习研究，并领导该系的 Agent Lab 实验室。他的研究方向包括深度强化学习和用于决策的深度学习。孙博士是人工智能复杂决策博士培训中心 (AI Centres for Doctoral Training Programme in Complex Decision-Making) 的招生委员会成员，并负责学生管理。在加入曼彻斯特大学之前，孙博士曾在微软剑桥研究院担任研究员，专注于基础模型研究；在牛津大学担任博士后研究员，专注于多智能体系统的研究工作。孙博士的研究成果在多个顶级机器学习和人工智能会议上获得认可，包括 NeurIPS、ICLR、KDD、AAAI、IJCAI 和 AAMAS。他还荣获 AAMAS 最佳论文奖和 Google Summer of Code 奖。