俞扬 - 南京大学 - 人工智能学院

个人简介

Recent highlights NeoRL: A near real-world offline RL benchmark An offfline RL competition on sales promotion ZOOpt: A Python package for derivative-free optimization. Introduction I received my Ph.D. degree in Computer Science from Nanjing University in 2011 (supervisor Prof. Zhi-Hua Zhou), and then joined the LAMDA Group (LAMDA Publications), in the Department of Computer Science and Technology of Nanjing University as an Assistant Researcher from 2011, and as an Associate Professor from 2014. I joined the School of Artificial Intelligence of Nanjing University as a Professor from 2019. My research interest is in reinforcement learning, and I am running a startup, Polixir.ai, to land reinforcement learning in real-life tasks. Research I am mainly focusing on reinforcement learning. Reinforcement learning searches for a policy of near-optimal decisions, by learning from environment interactions autonomously. Despite the fantastic future, reinforcement learning is still in its early infancy. Its potential has not been fully released in real-life tasks yet. Our team is trying in various aspects to improve reinforcement learning, including theoretical foundation, optimization, model structure, experience reuse, abstraction, model building, etc., heading toward sample-efficient methods for large-scale real-world applications. [ Publication list ] [ DBLP page ] [ Google Scholar page ] Codes ● GitHub: https://github.com/eyounx?tab=repositories ● LAMDA codes: http://www.lamda.nju.edu.cn/Data.ashx Teaching ● Tutorial of Artificial Intelligence (for undergraduate students of AI School. Fall, 2018) ● Advanced Machine Learning. (for graduate students. Fall, 2018) ● Advanced Machine Learning. (for graduate students. Fall, 2017) ● Artificial Intelligence. (for undergraduate students. Spring, 2015, 2016, 2017, 2018) ● Data Mining. (for M.Sc. students. Fall, 2014, 2013, 2012) ● Digital Image Processing. (for undergraduate students from Dept. Math., Spring, 2014, 2013) ● Introduction to Data Mining. (for undergraduate students. Spring, 2013, 2012)

近期论文

查看导师新发文章（温馨提示：请注意重名现象，建议点开原文通过作者单位确认）

● Real-world reinforcement learning aims at improving the sample efficiency of reinforcement learning and enabling real-world autonomous decision-making. ○ Adversarial environment model learning reconstructs environment models from offline data. Our work was the first adversarial approach for environment model learning, which solves the long-lasting compounding error issue of model-based RL for the first time. ○ Real world Experience reuse in reinforcement learning (with Qing Da, Chao Zhang, Zhi-Hua Zhou, etc.) Our studies design ways to accelerate reinforcement learning by resuing experiences, particularly, accumulated in simulators. ○ Reinforcement learning on StarCraft (with Zhen-Jia Pang, Ruo-Zhe Liu, etc.) Our studies try as efficiently as possible to learn a good playing policy for this extremely large-scale partial-observable real-time-strategy game. ● Derivative-free optimization (with Hong Qian, Yi-Qi Hu, etc.) aims at tackling optimization problems with complex structures, such as non-convex, non-differentiable, and non-continuous problems with many local optima. We developed classification-model-based derivative-free optimization, which not only serves as a theoretical framework but also an efficient optimization algorithm for local-Holder functions. The algorithms are collected in the toolbox ZOOpt for single and distributed optimization. ● Evolutionary Learning with Chao Qian, Ke Tang, Xin Yao, Zhi-Hua Zhou, etc.) ○ Approximation analysis & Pareto optimization: Our studies analyzed the goodness of solutions of evolutionary algorithms, and designed the Pareto optimization that shows as a powerful approximation tool for various subset selection problems. ○ Running time analysis of evolutionary optimization: We developed tools for analyzing the complexity of evolutionary algorithms, one of the most fundamental issues of evolutionary algorithms. Z.-H. Zhou, Y. Yu, C. Qian. Evoluionary Learning: Advances in Theories and Algorithms. Berlin: Springer, 2019. 周志华, 俞扬, 钱超. 演化学习: 理论与算法进展. 人民邮电出版社, 2021.