人工智能培训

搜索

机器学习论文:伪排练:在没有灾难遗忘的情况下实现深度强化学习(Pseudo-Rehearsal: Achieving Deep Reinforcement Le

[复制链接]
376156679 发表于 2018-12-7 11:39:18 | 显示全部楼层 |阅读模式
376156679 2018-12-7 11:39:18 38 0 显示全部楼层
机器学习论文:伪排练:在没有灾难遗忘的情况下实现深度强化学习(Pseudo-Rehearsal: Achieving Deep Reinforcement Learning without  Catastrophic Forgetting)神经网络可以在各种各样的任务上取得非凡的成果。然而,当他们试图顺序学习一些任务时,他们倾向于学习新任务,同时破坏性地忘记以前的任务。该问题的解决方案是伪排练,其涉及在排练代表先前任务的生成项目时学习新任务。我们认为将伪排练方法与生成网络配对是强化学习中这一问题的有效解决方案。我们的方法有效地学习了三个Atari 2600游戏,同时在所有三个游戏中保持了高于人类水平的表现,表现类似于从所有先前学习的任务中排练实例的网络。
Neural networks can achieve extraordinary results on a wide variety of tasks.However, when they attempt to sequentially learn a number of tasks, they tendto learn the new task while destructively forgetting previous tasks.Onesolution to this problem is pseudo-rehearsal, which involves learning the newtask while rehearsing generated items representative of previous task/s.Wedemonstrate that pairing pseudo-rehearsal methods with a generative network isan effective solution to this problem in reinforcement learning.Our methoditeratively learns three Atari 2600 games while retaining above human levelperformance on all three games, performing similar to a network which rehearsesreal examples from all previously learnt tasks.机器学习论文:伪排练:在没有灾难遗忘的情况下实现深度强化学习(Pseudo-Rehearsal: Achieving Deep Reinforcement Learning without  Catastrophic Forgetting) J5vr5V7VJnDvEk87.jpg
URL地址:https://arxiv.org/abs/1812.02464     ----pdf下载地址:https://arxiv.org/pdf/1812.02464    ----机器学习论文:伪排练:在没有灾难遗忘的情况下实现深度强化学习(Pseudo-Rehearsal: Achieving Deep Reinforcement Learning without  Catastrophic Forgetting)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

376156679当前离线
新手上路

查看:38 | 回复:0

快速回复 返回顶部 返回列表