Zeyu Zheng, a fourth-year PhD student in CSE, has been awarded a 2021 JP Morgan PhD Fellowship. Zeyu is advised by Toyota Professor of Artificial Intelligence Satinder Singh Baveja.
Zeyu’s long-term research goal is to build adaptive reinforcement learning (RL) agents that learn effectively and efficiently from their experience in real-world environments.
Despite its recent success in domains like Go and StarCraft, RL has yet to find a broad impact on real-world applications. One major challenge is its low sample efficiency: an RL agent usually needs millions or even billions of samples to solve a problem, which is impractical in most real-world scenarios. Consequently, practitioners often find themselves spending a lot of time and effort injecting their domain knowledge into the agent design or the environment design in order to improve the agent’s learning efficiency.
Through his research, Zeyu aims to reduce the burden on RL practitioners by automatically acquiring the knowledge that is typically injected by agent designers from experience. Specifically, he is exploring three different approaches:
1) meta-learning intrinsic rewards to reduce the workload of reward engineering.
2) discovering temporal credit assignment structures embedded in the environment to adapt the agent’s learning dynamics online.
3) generating auxiliary tasks via simple yet effective protocols to improve state representation learning.
His research has seen positive empirical results on improving data efficiency and demonstrates potential in bridging the gap between RL techniques and real-world problems.
Zeyu has previously been awarded a Rackham International Student Fellowship, in 2018.