读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning

DARLA:Improving Zero-Shot Transfer in Reinforcement Learning这篇文章发表于ICML2017,主要讲Reinforement learning算法在不同数据分布上的迁移。

鉴于实验室讲论文讨论的要求,将该论文做成了ppt,供大家分享。

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第1张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第2张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第3张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第4张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第5张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第6张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第7张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第8张图片


读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第9张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第10张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第11张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第12张图片

读论文-DARLA:Improving Zero-Shot Transfer in Reinforcement Learning_第13张图片

总结,这篇文章关键点在于使用DAE将real world的state转换为latent state。另外,DARLA不能称为严格意义的zero-shot,因为模拟环境与现实环境具有极大的相似性,致使在本质上Agent处于现实环境中训练(虽然表面上是虚拟环境训练)。不过,这文章的这种思想不错,能够在虚拟环境训练,并成功用于现实环境。

你可能感兴趣的:(Robot,learning)