Relay Generalization of Reinforcement Learning
- LecturerMr. Li-Cheng Lan (Computer Science, University of California, Los Angeles)
Host: Ti-Rong Wu - Time2022-12-15 (Thu.) 14:00 ~ 16:00
- LocationAuditorium108 at IIS old Building
Abstract
Generalization is critical for deploying reinforcement learning (RL) agents into real-world applications. A well-trained RL agent that can achieve high rewards under restricted settings may not be able to handle the enormous state space and complex environment variations in the real world. We propose, study, and improve the "relay-generalization" — a well-generalized agent that masters a task should be able to start from any “controllable” state in the same environment and still reach the goal. For example, a self-driving system may need to take over the control from humans in the middle of driving and must continue to drive the car safely. We show that the current RL methods are not general enough to pass our "relay-evaluation". We also propose a novel method called Self-Trajectory Augmentation (STA) which can reduce the failure rate dramatically.