首页 > AI前沿 > Learning from Relative Action Credit for Escaping Local Minima

Learning from Relative Action Credit for Escaping Local Minima

arXiv机器学习 2026-10-28 06:38 4 阅读 查看原文
首先,任务是把纯文本新闻正文排版成语义化的HTML片段。规则如下: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释。 2. 每个自然段落用

包裹,长文用

/

划分小节。 3. 列表内容用
  • ,引用用
    ,代码用
    ,重要句子可用。
    
    4. 保留原文所有信息,不得增删改内容。
    
    5. 不要输出//等文档级标签。
    
    正文内容是关于强化学习的:
    
    "Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data. In practice, on-policy methods such as PPO routinely stall on long-horizon, contact-rich tasks: dense reward shaping carries the policy through stages, yet stitching them together and discovering sparsely rewarded events rarely succeeds. We trace this failure to the missing action discrimination in state-value credit assignment: a state-value critic judges progress only by episode outcomes, so when the policy converges to a partial solution all outcomes coincide, and improvement stops because no action can prove better than another. We therefore extend PPO with an action conditioned value critic which, can compare the executed action with alternatives sampled at the same state and aid the learning signal. Combined with Generalized Advantage Estimation and a bound on exploration noise, we show how the extended algorithm escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails."
    
    我需要分析这个文本:
    
    - 它是一个连续的段落,没有明显的分节。
    
    - 但规则说"长文用

    /

    划分小节"。这里可能不算特别长,但为了符合规则,我应该考虑是否需要划分小节。 文本结构: - 第一句:介绍强化学习的潜力。 - 第二句:描述问题:PPO在长程任务中停滞。 - 第三句:解释原因:状态值信用分配中缺少动作判别。 - 第四句:提出解决方案:扩展PPO with action conditioned value critic。 - 第五句:结果:结合GAE和探索噪声界限,显示算法成功。 由于它只有一段,我可能不需要用

    或

    ,因为规则说"长文",但这里不算很长。规则说"长文用

    /

    划分小节",但如果没有明显的小节,我应该只用

    。 规则2:"每个自然段落用

    包裹,长文用

    /

    划分小节"。这个文本是一个自然段落,所以应该用

    包裹整个东西。 但为了安全,我应该检查是否有逻辑分节。文本中: - 开头:Reinforcement learning holds the promise... - 然后:In practice, on-policy methods... - 然后:We trace this failure... - 然后:We therefore extend PPO... - 然后:Combined with... 这可以视为一个小节,但可能不需要标题。规则说"长文",但这里只有几句话。我认为对于这个长度,直接用

    包裹整个段落即可。 规则4:保留原文所有信息,不得增删改内容。所以我不能添加任何标题。 规则3:提到列表、引用、代码等,但原文中没有这些元素。原文是纯叙述,没有列表、引用或代码。所以不需要用

      、
      或
      。
      
        重要句子可用。我需要判断哪些句子是重要的。
      
        例如:
      
        - "Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data." 这可能重要。
      
        - "We trace this failure to the missing action discrimination in state-value credit assignment" 可能重要。
      
        - "We therefore extend PPO with an action conditioned value critic" 可能重要。
      
        - "we show how the extended algorithm escapes the plateau and completes precise peg insertion" 可能重要。
      
        但规则说"重要句子可用",意思是如果句子重要,可以用标签包裹,但不是必须的。我可以选择应用。
      
        为了语义化,我应该只在关键点使用。
      
        现在,构建HTML:
      
        - 整个文本是一个段落,所以用

      包裹。 - 但规则说"每个自然段落用

      包裹",这里只有一个自然段落。 - 没有小节需要划分,因为没有明显的部分。 为了符合规则,我可以考虑将文本分成逻辑小节。 例如: - 介绍部分 - 问题描述 - 原因分析 - 解决方案 - 结果 但规则说"不得增删改内容",所以我不能添加标题;我只能基于原文结构。 原文是一个连贯的段落,没有分节。所以,我应该只用一个

      。 规则2:"长文用

      /

      划分小节"。这个文本不算很长,可能不需要。我可以假设它是短文。 为了安全,我会用

      包裹整个东西。 现在,关于重要句子:我可以用标记关键句子。 例如: - 第一句:重要,因为它介绍主题。 - "We trace this failure to the missing action discrimination" 重要。 - "We therefore extend PPO with an action conditioned value critic" 重要。 - 最后一句:重要。 但不要过度使用;只用于真正重要的。 在HTML中,用于强调重要性。 另外,规则3说"重要句子可用",所以是可选的。 我决定标记几个关键句子。 具体文本: "Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data." → 重要 "In practice, on-policy methods such as PPO routinely stall on long-horizon, contact-rich tasks: dense reward shaping carries the policy through stages, yet stitching them together and discovering sparsely rewarded events rarely succeeds." → 描述问题,可能重要 "We trace this failure to the missing action discrimination in state-value credit assignment: a state-value critic judges progress only by episode outcomes, so when the policy converges to a partial solution all outcomes coincide, and improvement stops because no action can prove better than another." → 核心原因,重要 "We therefore extend PPO with an action conditioned value critic which, can compare the executed action with alternatives sampled at the same state and aid the learning signal." → 解决方案,重要 "Combined with Generalized Advantage Estimation and a bound on exploration noise, we show how the extended algorithm escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails." → 结果,重要 但整个段落是连贯的,所以用

      包裹,内部用标记关键短语。 规则说"重要句子可用",意思是整个句子可以用包裹。 例如:

      Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data. In practice, ...

      但不要标记所有句子,只标记最重要的。 我选择标记: - 第一句 - "We trace this failure to the missing action discrimination in state-value credit assignment" - "We therefore extend PPO with an action conditioned value critic" - 最后一句 但"We trace..." 是句子的一部分。 为了精确,我应该用包裹整个关键句子。 列出关键句子: 1. Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data. 2. We trace this failure to the missing action discrimination in state-value credit assignment. 3. We therefore extend PPO with an action conditioned value critic which, can compare the executed action with alternatives sampled at the same state and aid the learning signal. 4. Combined with Generalized Advantage Estimation and a bound on exploration noise, we show how the extended algorithm escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails.