# News

• Two papers are accepted at NeurIPS 2020!
(1) provably sample-efficient RL for learning $Q$-value that “overcomes” the classical lower bound via low-rank matrix estimation.
(2) improved learning with imbalanced data by studying the value of labels in semi- and self-supervised manners.
• Work on harnessing the low-rank structure of $Q$-value for planning & deep RL is accepted as an oral presentation (1.8%) at ICLR 2020.
• Paper on non-asymptotic analysis of monte carlo tree search appears in ACM Sigmetrics 2020; we prove that the correct bounus term should be polynomial rather than the classical logarithmic one.
• Paper on private sequential learning is accepted at the journal Operations Research (preliminary version: COLT2018).

# Research

I am interested in both theoretical machine learning and modern deep learning applications.

#### Theory: novel mathematical models and theoretical guarantees for

• reinforcement learning algorithms
• private active learning

#### Applications: theory-inspired approaches in deep learning

• deep reinforcement learning
• class-imbalanced learning

# Selected Publications

(Google Scholar; theoretical work: alphabetical order)

Journal Papers and Preprints

Conference Papers

1. Previously circulated under the title “On Reinforcement Learning Using Monte Carlo Tree Search with Supervised Learning: Non-Asymptotic Analysis.” ^

# Work Experience

#### Tower Research Capital

Jun 2019 – Aug 2019 New York

#### Cubist Systematic Strategies, Point72 Asset Management

Jun 2018 – Aug 2018 New York

# Teaching

Over the years, I have TA’ed several graduate-level machine learning and optimization courses in the Department of Electrical Engineering and Computer Science at MIT.

##### 6.867 Machine Learning (Fall 2017 & Fall 2018)
• graduate-level introduction to the principles, techniques, and algorithms for modern machine learning.
##### 6.251 Introduction to Mathematical Programming (Spring 2017)
• graduate-level introduction to linear optimization and its extensions emphasizing both methodology and the underlying mathematical structures and geometrical ideas.
##### 6.231 Dynamic Programming and Stochastic Control (Fall 2016)
• graduate-level introduction to sequential decision-making via Markov decision processes