Research Papers 论文研究 9h ago Updated 5h ago 更新于 5小时前 35

When Do Options Help? Policy Necrosis and Redundant Coverage in Option-Critic When Do Options Help? Policy Necrosis and Redundant Coverage in Option-Critic

The learned termination rule in Option-Critic is functionally useless: when the termination test and option-picking policy observe identical values, it fires at every step, effectively collapsing to always-terminate behavior The authors introduce "policy necrosis," a phenomenon where sub-policies inside options barely explore and permanently lock onto their first seemingly-good action, affecting approximately 60% of states in a typical option Adding more options does not improve any individual o Option-Critic的终止规则在最大化回报的学习过程中实际贡献为零,当终止测试与策略读取相同值时会退化为"总是终止" 算法存在严重的探索不足问题,导致"政策坏死"(policy necrosis)现象,典型选项中约3/5的状态陷入坏死 增加选项数量并不能改善单个选项的性能,而是通过降低所有选项在同一状态同时失败的概率(从59%降至4%)来提升整体表现

50
Hot 热度
50
Quality 质量
50
Impact 影响力

Analysis 深度分析

TL;DR

  • The learned termination rule in Option-Critic is functionally useless: when the termination test and option-picking policy observe identical values, it fires at every step, effectively collapsing to always-terminate behavior
  • The authors introduce "policy necrosis," a phenomenon where sub-policies inside options barely explore and permanently lock onto their first seemingly-good action, affecting approximately 60% of states in a typical option
  • Adding more options does not improve any individual option's performance; instead, it reduces the joint probability of all options failing in the same state from 59% to 4%, which drives the observed performance gains
  • Forcing termination at every step preserves the option-count performance curve, confirming the termination mechanism contributes nothing meaningful
  • Restoring exploration within options repairs necrotic states, enabling a single option to solve the task

Why It Matters

This paper provides a critical deconstruction of Option-Critic, one of the foundational architectures for hierarchical reinforcement learning, revealing that its celebrated performance improvements from adding options stem from redundant coverage rather than genuine sub-policy learning. For practitioners building or extending hierarchical RL systems, these findings challenge assumptions about the value of option count and highlight exploration as the true bottleneck.

Technical Details

  • Termination rule analysis: The paper proves that when the termination test and the option-selection policy read the same value function, the termination condition fires at every timestep, making the learned rule equivalent to always terminating. In cases where the policy explores but the termination test does not (as in standard Option-Critic), the termination rule can actively block exploration, suffering Omega(T) regret compared to O(log T) for always-terminating.
  • Policy necrosis: The authors define and detect a state-level phenomenon where the policy inside an option ceases exploration almost entirely, locking onto the first action with favorable initial estimates. They quantify this by finding three-fifths of states exhibit necrotic behavior in a typical option.
  • Redundant coverage mechanism: Performance scaling with option count is explained through a probabilistic lens: additional options do not make any single option better, but they reduce the probability that all options fail simultaneously in a given state, from 59% down to 4%.
  • Empirical validation: The authors demonstrate that restoring exploration within options repairs necrotic states, allowing even a single option to solve tasks that previously required multiple options.

Industry Insight

  • Hierarchical RL practitioners should prioritize exploration mechanisms within sub-policies over simply increasing option count; the marginal gains from additional options appear to come from coverage redundancy rather than genuine skill diversification.
  • The policy necrosis phenomenon likely generalizes beyond Option-Critic to other hierarchical RL frameworks that use fixed-duration or internally-terminated sub-policies, suggesting a need for built-in exploration guarantees at the sub-policy level.
  • The finding that termination rules can actively harm exploration warrants re-examination of termination learning objectives in existing hierarchical RL systems, particularly in safety-critical applications where option switching behavior directly impacts system reliability.

TL;DR

  • Option-Critic的终止规则在最大化回报的学习过程中实际贡献为零,当终止测试与策略读取相同值时会退化为"总是终止"
  • 算法存在严重的探索不足问题,导致"政策坏死"(policy necrosis)现象,典型选项中约3/5的状态陷入坏死
  • 增加选项数量并不能改善单个选项的性能,而是通过降低所有选项在同一状态同时失败的概率(从59%降至4%)来提升整体表现

为什么值得看

这篇论文对Option-Critic这一经典分层强化学习框架进行了深刻的理论剖析,揭示了其性能提升背后的真实机制,对理解分层RL的局限性具有重要价值。研究提出的"政策坏死"概念和状态级检测方法,为改进分层策略学习提供了新的分析工具。

技术解析

  • Option-Critic学习选项(子策略)和终止规则,其核心假设是性能随选项数量增加而提升,但论文通过理论和实验证明这一结论的机制被误解
  • 终止规则的理论分析显示:当终止测试与策略读取相同值时,测试每一步都会触发,等效于总是终止;而当策略探索而测试不探索时,规则可能阻碍探索,导致Ω(T) regret
  • "政策坏死"现象:选项内部策略几乎不探索,状态锁定在第一个看似好的动作上不再更新,论文提出状态级测试方法并发现典型选项中3/5的状态存在坏死
  • 恢复探索能力可修复坏死状态,使单个选项即可完成任务,证明额外选项的价值在于冗余覆盖而非性能提升

行业启示

  • 分层强化学习中的"更多选项=更好性能"假设需要重新审视,性能提升可能源于概率冗余而非策略质量改进,设计时应关注探索机制的有效性
  • "政策坏死"现象在分层RL中可能普遍存在,建议在选项学习框架中强制引入探索机制,避免状态过早锁定
  • 评估分层RL系统时,应区分"单个选项能力提升"与"多选项覆盖概率提升",后者可能掩盖单个策略的缺陷,需建立更精细的评估指标

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。