Hello! I am a PhD student at Boston University, advised by Prof. Yannis Paschalidis and Prof. Ashok Cutkosky.
I am broadly interested in the theoretical aspects of machine learning, optimization and control theory. Specifically, I work on adaptive online learning, i.e., designing online decision making algorithms that optimally exploit problem structures.
Prior to BU, I studied mechanical engineering at Tsinghua University.
Email address: zhiyuz [at] bu (dot) edu
Google Scholar Github LinkedIn
Research
The latest versions are available on arXiv. The code of my published papers is on Github.
- Optimal Parameter-free Online Learning with Switching Cost
ZZ, Ashok Cutkosky, Ioannis Paschalidis.
arXiv preprint.We design parameter-free algorithms in the presence of switching costs. Using a dual space scaling strategy, our algorithm effectively balances these two opposite considerations, improving the existing regret bound to the optimal rate.
- PDE-Based Optimal Strategy for Unconstrained Online Learning
ZZ, Ashok Cutkosky, Ioannis Paschalidis.
ICML 2022.We design a parameter-free online learning algorithm by solving a PDE. Our framework requires less guessing than prior approaches, and the obtained regret bound achieves the optimal leading constant for the first time.
- Adversarial Tracking Control via Strongly Adaptive Online Learning with Memory
ZZ, Ashok Cutkosky, Ioannis Paschalidis.
AISTATS 2022.We reveal a connection between two separate notions of “tracking” in online learning and linear control. Through several new results in adaptive online learning, we propose a controller that tracks an adversarially generated state sequence with a strong guarantee.
- Provable Hierarchical Imitation Learning via EM
ZZ, Ioannis Paschalidis.
AISTATS 2021. Also presented at ICML 2020 Workshop.We use the framework of Expectation-Maximization to analyze the learning of a hierarchical policy from expert demonstrations.
Service
Reviewer for ICML, NeurIPS, AISTATS.
Recognitions: Top reviewer (10%) at AISTATS 2022.