强化学习：从经验中学习

Reinforcement Learning: Learning from Experience

理查德·萨顿 Richard Sutton · Stanford · 2026-03-16 · 约 34 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

Rich Sutton 探讨强化学习的历史与力量，这是一种无需教师、通过试错从经验中学习的机器学习方法。

Rich Sutton discusses the history and power of reinforcement learning, a machine learning approach that learns from trial and error without a teacher.

要点 · TL;DR

• 强化学习通过试错从经验中学习以实现目标。
  Reinforcement learning learns from trial and error to achieve goals.
• 时序差分学习是该领域最好的贡献。
  Temporal difference learning is the best contribution of the field.
• 学习适合学习的表征是首要开放问题。
  Learning good representations for learning is the top open problem.

核心观点 · Key points

• 强化学习是通过试错从经验中学习以实现目标。
  Reinforcement learning is learning from experience by trial and error to achieve a goal.
• 时序差分学习是我做过的最好的事。
  Temporal difference learning is the best thing I ever did.
• 连接主义方法战胜符号主义方法是一个大惊喜。
  The victory of connectionist methods over symbolic methods is a big surprise.
• 深度学习没有学到有利于学习的表征。
  Deep learning does not learn representations that are good for learning.
• 学习有利于学习的表征是首要的开放问题。
  Learning representations that are good for learning is the number one open problem.
• 使用学到的模型进行规划是另一个重要的开放问题。
  Planning with a learned model is another important open problem.

反共识 · Contrarian takes

• 我一生中从未做过任何直接有用的事情。
  I have never done anything directly useful in my whole life.
• 神经网络不是神经的，应该被称为连接主义系统。
  Neural networks are not neural; they should be called connectionist systems.
• 监督学习的进展没有我预期的那么远。
  Supervised learning hasn't progressed as far as I would have expected.
• 我们不知道如何学习好的表征，而且我们已经不再谈论它了。
  We don't know how to learn good representations, and we've stopped talking about it.
• 炒作和行业扭曲了科学；AI 寒冬甚至可能是好事。
  The hype and industry warp the science; an AI winter might even be good.
• 可能会出现对 AI 的歧视，但最终我们应该接受它们作为心智。
  There may be discrimination against AIs, but eventually we should accept them as minds.

本期章节 · Chapters（共 10）

• 强化学习简介 Introduction to Reinforcement Learning
• 时序差分学习 Temporal Difference Learning
• 强化学习的成功案例 Successes of Reinforcement Learning
• 该领域最大惊喜 Biggest Surprises in the Field
• 监督学习与表征学习之意外 Surprise about supervised learning and representation learning
• 开放问题与博士论文选题 Open problems and PhD thesis topics
• AI 与神经科学的联系 Connection between AI and neuroscience
• 机器学习与神经科学的交叉性 Interdisciplinary nature of ML and neuroscience
• AI 的未来及学生建议 Future of AI and advice to students
• 结束语 Closing remarks

阅读全文双语转录 →