Page History: Yang Yu @ NJUCS
Compare Page Revisions
Page Revision: 2015/05/08 21:49
| Yang Yu (Y. Yu) Can be pronounced as "young you" Ph.D., Associate Professor LAMDA Group Department of Computer Science National Key Laboratory for Novel Software Technology Nanjing University
Office: 311, Computer Science Building, Xianlin Campus email: yuy@lamda.nju.edu.cn, eyounx@gmail.com, yuy@nju.edu.cn (the last email is both symbolic and problematic) | |
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), Department of Computer Science and Technology of Nanjing University as an assistant researcher in 2011, and as an associate professor in 2014.
I am interested in artificial intelligence, particularly, applying theoretical-grounded evolutionary optimization to solve machine learning problems. Our research work has been published in international journals and conferences including Artificial Intelligence, IJCAI, KDD, etc., and granted several awards such as Grand Champion (with other LAMDA members) of PAKDD'06 Data Mining Competition, the best paper award of PAKDD'08, the best theory paper award of GECCO'11. My dissertation was granted honours including China National Outstanding Doctoral Dissertation Award in 2013 and China Computer Federation Outstanding Doctoral Dissertation Award in 2011.
(
Detailed CV)
Recent Update
- Invited Talk: On Evolutionary Machine Learning (PDF, 5.75MB) (The 12th China Workshop on Machine Learning and Applications, Nov. 2014, Xi'an, China)
- Invited Talk: The Sampling-and-Classification Framework (The 10th NICaiA Workshop on Nature Inspired Computation and Its Applications, Nov. 2014, Hefei, China)
- Co-Chair: ECOLE'14 (1st Chinese Workshop on Evolutionary Computation and Learning, July 2014, Hefei, China)
- Invited Talk: Analyzing Evolutionary Optimization in Noisy Environments (The 8th NICaiA Workshop on Nature Inspired Computation and Its Applications, Oct. 2013, Nanjing, China)
- Our tutorial "An Introduction on Evolutionary Optimization: Recent Theoretical and Practical Advances" (with Ke Tang, Xin Yao and Zhi-Hua Zhou) has been accepted by IJCAI'13 (IJCAI Schedule) (Tutorial Webpage)
Research
I am interested in investigating towards the goal of
artificial intelligence. As conceived by Alan Turing, one possible way of building an
intelligent machine is to evolve learning machines, which now drops into multiple subfields of artificial intelligence, especially
machine learning and
evolutionary computation. Currently I am focusing on the subfields:
- Theoretical foundation of evolutionary algorithms (my related publications): to understand and reveal the optimization power of evolutionary algorithms, in order to solve problems in machine learning;
- Ensemble learning (my related publications): to achieve the strongest generalization ability from finite samples;
- Reinforcement learning (my related publications): to be intelligent to survive in environments autonomously.
(Evolutionary Machine Learning)
- Ensemble pruning: (with Chao Qian and Zhi-Hua Zhou)
Ensemble pruning can drastically improve the generalization performance of an ensemble while saving storage and prediction time costs. Search for the best subset among the base learners is its key problem, which is unfortunately NP-hard. Incorporating evolutionary optimization, we propose the Pareto ensemble pruning (PEP) method to tackle this problem in (PDF-aaai15). Both theoretical and empirical studies show the advantages of PEP to the state-of-the-art approaches.
(Evolutionary Optimization)
- The sampling-and-learning (SAL) framework: (with Hong Qian)
Most of evolutionary algorithms share a common algorithm structure, which has been argued to involve the sampling and the model building stages. In (PDF-cec14), we developed the sampling-and-learning (SAL) framework to capture the essence of these algorithms, and analyzed its approximation performance under a probability. Focusing on the sampling-and-classification (SAC) algorithms, which specify the learning sub-routine of SAL as a binary classification, we dervied a general performance bound, with the help of the learning theory. We also disclosed conditions under which SAC algorithms can achieve polynomial and super-polynomial speedup to random search.
- Approximation ability of evolutionary optimization: (with Chao Qian, Xin Yao and Zhi-Hua Zhou)
Evolutionary algorithms are most commonly used to obtain good-enough solutions rather than optimal solutions in practice, which relates to their approximation ability.- We developed the SEIP framework (PDF-aij12) to characterize the approximation ability of evolutionary algorithms, and showed that they can achieve the currently best-achievable approximation ratio for the k-set cover problem, which also reveals the advantage of evolutionary algorithms over the well known greedy algorithm.
- We disclosed that crossover operator can help fulfill the Pareto front in multi-objective optimization tasks. As a consequence, on the bi-objective Minimum Spanning Tree problem, we show that a multi-objective EA with a crossover operator improves the running time from that without the crossover for achieving a 2-approximate solution (PDF-aij13).
- General approaches to running time analysis of evolutionary optimization: (with Chao Qian and Zhi-Hua Zhou)
The running time, or the computational complexity, of evolutionary algorithms and other metaheuristic algorithms is one of the most important theoretical issues for understanding these algorithms.- We developed the switch analysis (PDF-tec15) for running time analysis of evolutionary algorithms. It compares two algorithm processes (a.k.a. two Markov chains), leading to the relationship of the expected running time of the two processes.
- Previous analysis approaches, including the drift analysis and the fitness level method are reducible to switch analysis, revealing the power of switch analysis (PDF-tec15).
- We also proved that our convergence-based approach is also reducible to switch analysis (PDF-cec15).
- Using switch analysis to compare problem instances, we can identify the hardest and the easiest problem instances for a metaheuristic algorithm in a problem class (PDF-tec15, PDF-ppsn12), demonstrating the ability of switch analysis for problem class-wise analysis.
- We applied switch analysis to analyzing crossover operators (tr11, PDF-ppsn10) that can lead to tight running time bounds.
- We developed the convergence-based approach (PDF-aij08) that can be applied to obtain running time bounds of a large range of metaheuristic algorithms. We applied this tool to disclose that infeasible solutions in constrained optimization may not be useless (PDF-cec08).
(Machine Learning)
- Functional representation for nonlinear reinforcement learning: (with Yu Qian, Qing Da and Zhi-Hua Zhou)
Reinforcement learning seeks for a policy that receives the highest total reward from its environement. Real-world applications are often so complex that complex policies are appealing. Functional representation, known as boosting techniques widely used in supervised learning, is a powerful tool to approximate complex functions but has been little investigated in reinforcement learning.- By functional representation, a policy can be a sum of many basis functions, which causes a high time cost especially for reinforcement learning. We tackle this time cost barrier by the napping mechanism, which results a significant improvement in time as well as a potential improvement in total reward (PDF-aamas14 / code)
- The role of diversity in ensemble learning: (with Nan Li, Yu-Feng Li and Zhi-Hua Zhou)
Ensemble learning is a machine learning paradigm that achieves the state-of-the-art performance. Diversity was believed to be a key to a good performance of an ensemble approach, which, however, previously served only as a heuristic idea.- By the diversity regularized machine (PDF-ijcai11 / code), we showed that diversity plays a role of regularization as in popular statistical learning approaches.
- We proved that diversity defined on hypothesis output space plays a role of regularization, and proposed the diversity regularized ensemble pruning to prune Bagging classifiers. (PDF-ecml12 / code)
Codes
(
My Goolge Scholar Citations)
Teaching
- Artificial Intelligence. (for undergraduate students. Spring, 2015) >>>Course Page>>>
- Data Mining. (for M.Sc. students. Fall, 2014)
- Digital Image Processing. (for undergraduate students from Dept. Math., Spring, 2014)
- Data Mining. (for M.Sc. students. Fall, 2013)
- Introduction to Data Mining. (for undergraduate students. Spring, 2013)
- Digital Image Processing. (for undergraduate students. Spring, 2013)
- Data Mining. (for M.Sc. students. Fall, 2012)
- Introduction to Data Mining. (for undergraduate students. Spring, 2012)
Students
Recuit >>>Current masters students:
2012:
Yu Qian
2013:
Hong Qian,
Jian Ma
2014:
Peng-Fei Hou,
Jing-Wen YangPast co-supervised masters:
Graduated in 2014: Xun-Li Guo (now in Alibaba Co.), Song Shuan (now in 360 Co.)
Mail: National Key Laboratory for Novel Software Technology, Nanjing University, Xianlin Campus Mailbox 603, 163 Xianlin Avenue, Qixia District, Nanjing 210023, China
(In Chinese:) 南京市栖霞区仙林大道163号,南京大学仙林校区603信箱,软件新技术国家重点实验室,210023。