Tian Xu

Ph.D. student, School of Artificial Intelligence, Nanjing University [Google Scholar] [CV] Email: xut@lamda.nju.edu.cn

About me

Welcome to my homepage! I am a Ph.D. student at Nanjing University (NJU), advised by Prof. Yang Yu.

My research interests lie in reinforcement learning theory and algorithms.

I visited The Chinese University of Hong Kong, Shenzhen (CUHKSZ) from June to September in 2021, where I was fortunate to be supervised by Prof. Zhi-Quan (Tom) Luo.

Selected Work

*: indicating equal contribution.

• Provably and Practically Efficient Adversarial Imitation Learning with General Function Approximation.
  Tian Xu*, Zhilong Zhang*, Ruishuo Chen, Yihao Sun, Yang Yu
  In Advances in Neural Information Processing Systems 38 (NeurIPS), 2024.
  (This work proposes a new AIL algorithm OPT-AIL. In theoy, OPT-AIL is the first provably efficient AIL algorithm with general function approximation.
  In practice, OPT-AIL achieves SOTA performance on a variety of tasks.)

• Imitation Learning from Imperfection: Theoretical Justifications and Algorithms.
  Ziniu Li*, Tian Xu*, Zeyu Qin, Yang Yu, Zhi-Quan Luo
  In Advances in Neural Information Process System 36 (NeurIPS, spotlight), 2023.
  (This work proposes a data-selection-based method ISW-BC to address the distribution shift issue in IL with imperfect demonstrations.
  We prove that ISW-BC is robust to OOD samples and enjoys a small imitation gap bound.)

• Provably Efficient Adversarial Imitation Learning with Unknown Transitions
  Tian Xu*, Ziniu Li*, Yang Yu, Zhi-Quan Luo
  In Proceedings of the 39th Conference on Uncertainty in Artificial Intelligence (UAI, oral presentation, acceptance rate < 3% ), 2023
  (This work proposes an “exploration-estimation-optimizatoin” error decomposition analysis for AIL in the unknown transition setting.
  Based on this analysis, we derive a new method MB-TAIL, which achieves the optimal expert sample complexity and better interaction complexity.)

• Understanding Adversarial Imitation Learning in Small Sample Regime: A Stage-coupled Analysis
  Tian Xu*, Ziniu Li* , Yang Yu, Zhi-Quan Luo
  arXiv:2208.01899
  (This work provides the first horizon-free imitation gap bound of adversarial imitation learning (AIL),
  which answers the open question –why does AIL outperform BC by a wide margin, particularly in the low data regime ? –
  raised in CoRL 2019 best paper by Ghasemipour et al.)

• Rethinking ValueDice: Does It Really Improve Performance?
  Ziniu Li*, Tian Xu*, Yang Yu, Zhi-Quan Luo
  In Proceedings of the 10th International Conference on Learning Representations (ICLR) (Blog Track), 2022
  (This work demonstrates the first reduction of offline adversarial imitation learning (AIL) to BC, implying that AIL does not outperform BC in the offline setting.)

• Error Bounds of Imitating Policies and Environments
  Tian Xu*, Ziniu Li* , Yang Yu
  In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020
  (This work presents a general analysis framework for imitation learning algorithms, upon which we first prove that
  GAIL-type methods can overcome the compounding errors issue of BC in both imitating policies and environments.)

Service

Reviewer

NeurIPS (2022-), ICML (2022-), ICLR (2024-), AISTATS (2024-), UAI (2021-), RLC (2024-), EWRL (2022-).

Award

• [2020-10] National Scholarship for Graduates.

• [2020-8] Second winner on KDD CUP 2020 Reinforcement Learning Competition Track.

• [2016-9, 2017-9] National Scholarship for Undergraduates.