Ryan Zhiyu An

Postdoctoral Scholar

Department of Computer Science and Engineering

School of Engineering

University of California, Merced

Address: #230V, SE2, 5200 N. Lake Road, Merced, CA 95343

Email: zan7 AT ucmerced.edu

[Google Scholar]

Biography

Zhiyu An is a postdoctoral scholar at University of California, Merced. He received his Ph.D. in Electrical Engineering and Computer Science at University of California, Merced. He received his B.A. in Computer Science at UC Berkeley.

Academic Services

Program Committee Member
- AAAI '26
- ACM BuildSys '25
Conference Reviewer
- IROS '25
- SenSys '24
- CIKM '24 TRAI Workshop
Journal Reviewer
- IEEE Transactions on Network Science and Engineering (T-NSE), '24 ~
- IEEE Internet of Things Journal (IoT-J), '24 ~

Research Overview

My research is driven by the goal of building safe and responsible AI agents to assist or act on humans' behalf in complex, real-world environments. My research tackles this through alignment - aligning, for instance, the trade-off between aggressive, high-gain control policies for climate-controlled environments and the conservatism under weather prediction uncertainty. Aligning learning-based AI agents with human values poses fundamental challenges compared to settings where policy-learning excelled, i.e., simplistic reward functions. First, human preferences such as the mentioned trade-offs can vary significantly, and conditioning learning-based agents on different preferences is non-trivial. Second, we need techniques to verify the agent's alignment in out-of-distribution tasks, which is often over an intractably large task space. Third, there is a fundamental lack of data on human values in different tasks, while modeling human values and finding collective actions is non-trivial.

To tackle these challenges, I focus on safe model-based reinforcement learning through epistemic uncertainty estimation, time-series modeling, and policy distillation. I have applied the above techniques to improve the robustness and Pareto frontiers of indoor climate-control systems and agricultural aquifer recharge (BuildSys'23, DAC'24, KDD'24, ICLR'24). I leverage techniques such as Langevin dynamics sampling for disentangling epistemic and aleatoric uncertainties, which advances theoretical understanding of how AI systems can reason about different sources of uncertainty (L4DC'25). As AI systems become more sophisticated, I am expanding this foundation to address fundamental questions of AI alignment through moral decision frameworks and reasoning-level reinforcement learning for LLM agents.

Selected Publications

[L4DC] Zhiyu An, Zhibo Hou, Wan Du (2025). Disentangling Uncertainties by Learning Compressed Data Representation.

7th Annual Learning for Dynamics & Control Conference [Paper] [Code]

[HICSS] Xianzhong Ding, Wanshi Hong, Zhiyu An, Bin Wang, Wan Du (2025). Deepot: Parking Lot Identification Using Low-Resolution Satellite Imagery

The 58th Hawaii International Conference on System Sciences 🏆 Best Paper Award [Paper]

[IEEE IoT-J] Xianzhong Ding*, Zhiyu An*, Arya Rathee, Wan Du (2025). A Safe and Data-efficient Model-based Reinforcement Learning System for HVAC Control

IEEE Internet of Things Journal (* denotes equal contribution) [Paper]

[SIGIR Workshop] Zhiyu An, Xianzhong Ding, Yen-Chun Fu, Cheng-Chung Chu, Yan Li, Wan Du (2025). Golden-Retriever: High-Fidelity Agentic Retrieval Augmented Generation for Industrial Knowledge Base.

SIGIR workshop on Robust Information Retrieval [Paper] Work done while interning at Western Digital.

[ICLR] Zhiyu An, Xianzhong Ding, Wan Du (2024). Reward Bound for Behavioral Guarantee of Model-Based Planning Agents.

Tiny Paper Track, International Conference on Learning Representations (Invited to Present) [Paper] [Open Review]

[DAC] Zhiyu An, Xianzhong Ding, Wan Du (2024). Go Beyond Black-box Policies: Rethinking the Design of Learning Agent for Interpretable and Verifiable HVAC Control.

The 61st ACM/IEEE Design Automation Conference [Paper] [Code] [Slides] [Presentation Recording] [Poster]

[KDD] Yuning Chen, Kang Yang, Zhiyu An, Brady Holder, Luke Paloutzian, Khaled M. Bali, Wan Du (2024). MARLP: Time-series Forecasting Control for Agricultural Managed Aquifer Recharge

The 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining [Paper]

[BuildSys] Zhiyu An, Xianzhong Ding, Arya Rathee, Wan Du (2023). CLUE: Safe Model-Based RL HVAC Control Using Epistemic Uncertainty Estimation.

ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation 🏆 Best Paper Award Runner-Up [Paper] [Code] [Slides]

[SenSys Poster] Zhiyu An, Xianzhong Ding, Wan Du (2023). Data Efficient HVAC Control using Gaussian Process-based Reinforcement Learning

The ACM Conference on Embedded Networked Sensor Systems [Abstract]

Page updated

Google Sites

Report abuse