Reinforcement learning
Editable XRL
CoChoice makes an RL agent’s actions, predictions, and values editable through Markov trees. Corrections guide continued learning, helping people and agents coordinate their roles.
Hi, I’m Haoyang. 陈昊阳
AI that people can
understand and shape.
I design interactive systems that help people bring their knowledge and intentions into AI. My work turns explanations into ways to revise models, helps people coordinate with learning agents, and gives users control over the context of LLM conversations. I combine human–computer interaction and interpretable machine learning to explore how people and AI can better understand and adapt to one another.
I am a PhD student in the Ubicomp Lab at the National University of Singapore, advised by Prof. Brian Lim. Previously, I was a research assistant at COOLA Lab, Southeast University, working with Prof. Yan Lyu and Prof. Wanyuan Wang.
Publications
- arXiv 2026
-
TreePIN²: Tree-Prior Initialized Neural Networks
Y Wang, Q Sun, H Chen* - CHI'26
-
Editable XAI: Toward Bidirectional Human-AI Alignment with Co-Editable Explanations of Interpretable Attributes
H Chen, J Bai, T Fang, B Lim - AAAI'24
-
i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance
H Chen, P Sun, Q Song, W Wang, W Wu, W Zhang, G Gao, Y Lyu - AIAHPC'22
-
Multi-agent Reinforcement Learning for Fleet Management: A Survey
H Chen, Z Li, X Yao
* Corresponding author
Research projects
LLM sensemaking
Trellis
An interactive graph workspace for LLM-assisted sensemaking. People can branch into alternatives, merge insights, and reorganize conversations while controlling the context available to the model.
Model initialization
TreePIN²
Translates a fitted decision tree’s split rules, decision paths, and leaf predictions into neural network initialization for classification and regression. Training data calibrates split softness, and all weights and biases remain trainable.
Editable explanations
Editable XAI
CHI 2026CoExplain lets people revise explanations through rules on interpretable attributes. A study with 43 participants found improvements in understanding and alignment compared with read-only explanations.
Transportation
i-Rebalance
AAAI 2024Personalizes ride-hailing vehicle repositioning using driver preferences, response modeling, and reinforcement learning. A field study with 106 professional drivers informed the design.
Medical decision support
Interpretable Medical Decision Support
A visual programming toolkit for customizable diagnostic models. Doctors can express their preferred examination sequences and thresholds in an interpretable model.
Accessible interaction
Ear Motion for VR
An ear-motion input system for accessible interaction with VR devices. A study with 15 volunteers found the approach effective and easy to use.
Teaching
National University of Singapore
- CS4246/5446Reinforcement Learning and Sequential Decision Making AY2026/2027 · Sem 1
- CP2107Independent Introduction to CS Research (Odyssey) Summer 2026 · May–Aug
- CS3263Foundations of Artificial Intelligence AY2025/2026 · Sem 2
- CS3244Machine Learning AY2025/2026 · Sem 1
Service
- ACM Conference on Human Factors in Computing Systems (CHI) Reviewer · 2026, 2027
- AAAI Conference on Artificial Intelligence (AAAI) Reviewer · 2027
- IEEE Transactions on Intelligent Transportation Systems Reviewer · 2025