Jiamin He

I recently received my Computer Science M.Sc. (Thesis) degree at the University of Alberta. I am supervised by Rupam Mahmood. My current research interests lie in artificial intelligence, especially reinforcement learning.

I got my Bachelor's degree in Information and Computing Science at Sun Yat-sen University under the supervision of Hankz Hankui Zhuo. I was previously a research intern at Machine Intelligence Group at Tsinghua University led by Chongjie Zhang.

email  /  GitHub  /  blog

Loosely Consistent Emphatic Temporal-Difference Learning.
Jiamin He, Fengdi Che, Wan Yi, A. Rupam Mahmood.
Conference on Uncertainty in Artificial Intelligence (UAI), 2023.
paper | code

The Emphatic Approach to Average-Reward Policy Evaluation.
Jiamin He, Wan Yi, A. Rupam Mahmood.
Deep Reinforcement Learning Workshop at NeurIPS, 2022.
paper

Episodic Multi-agent Reinforcement Learning with Curiosity-Driven Exploration.
Lulu Zheng*, Jiarui Chen*, Jianhao Wang, Jiamin He, Yujing Hu, Yingfeng Chen, Changjie Fan, Yang Gao, Chongjie Zhang.
Conference on Neural Information Processing Systems (NeurIPS), 2021.
paper | code

Hindsight Multi-agent Credit Assignment
Vanilla policy gradient methods for MARL suffer from both variance and lack of counterfactuals, making the learning process data inefficient and unstable. We proposed a unified credit assignment framework to simultaneously assign credit over time and agents to address these issues. We proposed the Hindsight Multi-Agent Credit Assignment (HMACA), which uses both hindsight likelihood and counterfactual reasoning to assign credit over timesteps and agents.
report

Emergent Tool Use in Multi-agent Reinforcement Learning
We used PPO and adversarial training to train two teams of agents in a mixed cooperative-competitive asymmetric game environment we developed. During training, the agents learned to use ladders to cross the walls or block the adversaries, and surprisingly, they also learned to use ladders to speed up even if there were no walls or adversaries nearby. To make learning more efficient, we also adopted Hindsight Credit Assignment to address the temporal credit assignment.
demo

Sparkle Planning Challenge 2019 Entry: SYSU-Planner
We combined two planners into a two-phased planner that performs a fast best-first width search followed by a refinement hill-climbing search and obtained better overall performance. We achieved state-of-the-art of classical planning in 1/3 of the tested domains and outperformed all previous methods in a few domains. Our planner is in third place on the leaderboard and fourth place in the final result.
description | leaderboard

Stolen from Jon Barron