面向强化学习中更高样本效率的数据增强方法
人工智能
2019-11-18 v3 机器人学
摘要
深度强化学习(DRL)是一种有前景的自适应机器人控制方法,但其当前在机器人学中的应用受到高样本需求的阻碍。我们提出两种用于 DRL 的新颖数据增强技术,以更有效地复用观测数据。第一种称为万花筒经验回放(Kaleidoscope Experience Replay),利用反射对称性;第二种称为目标增强经验回放(Goal-augmented Experience Replay),利用宽松的目标定义。我们的初步实验结果显示学习速度大幅提升。
引用
@article{arxiv.1910.09959,
title = {Towards More Sample Efficiency in Reinforcement Learning with Data Augmentation},
author = {Yijiong Lin and Jiancong Huang and Matthieu Zimmer and Juan Rojas and Paul Weng},
journal= {arXiv preprint arXiv:1910.09959},
year = {2019}
}
备注
NeurIPS 2019 Workshop on Robot Learning: Control and Interaction in the Real World (accepted after double-blind peer review). arXiv admin note: substantial text overlap with arXiv:1909.10707