principles_of_deep_rl.pdf
40墨值下载
Principles of Deep RL
David Silver
Principle #1: Evaluation Drives Progress
Objective, quantitative evaluation drives progress:
● The choice of evaluation metric determines the direction of progress
● Arguably the most important single decision in the course of a project
Leaderboard-driven research:
● Be sure the evaluation metric corresponds closely to the end goal
● Avoid subjective evaluation (e.g. human inspection)
Hypothesis-driven research:
● Formulate a hypothesis:
○ “Double-Q learning outperforms Q-learning because it reduces upward bias”
● Verify hypothesis under a broad range of conditions
● Compare like-for-like not against existing state-of-the-art
● Seek understanding rather than leaderboard performance
of 11
40墨值下载
【版权声明】本文为墨天轮用户原创内容，转载时必须标注文档的来源（墨天轮），文档链接，文档作者等基本信息，否则作者和墨天轮有权追究责任。如果您发现墨天轮中有涉嫌抄袭或者侵权的内容，欢迎发送邮件至：contact@modb.pro进行举报，并提供相关证据，一经查实，墨天轮将立刻删除相关内容。
下载排行榜
评论