Yunzhe Tian, Yike Li, Yingxiao Xiang, Wenjia Niu, Endong Tong, and Jiqiang Liu (Beijing Jiaotong University)
Robust reinforcement learning has been a challenging problem due to always unknown differences between real and training environment. Existing efforts approached the problem through performing random environmental perturbations in learning process. However, one can not guarantee perturbation is positive. Bad ones might bring failures to reinforcement learning. Therefore, in this paper, we propose to utilize GAN to dynamically generate progressive perturbations at each epoch and realize curricular policy learning. Demo we implemented in unmanned CarRacing game validates the effectiveness.