引导式测验 · 硬 等级

强化学习测验

将强化学习概念应用于探索、折扣回报、延迟奖励、模拟限制和代理奖励目标的失败。

相关引导路径强化学习
问题 1 的 8

仓库机器人尝试一条陌生的过道，以了解是否可以缩短交货时间。这是探索与利用权衡的哪一方？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. 仓库机器人尝试一条陌生的过道，以了解是否可以缩短交货时间。这是探索与利用权衡的哪一方？
    • Exploration: testing an action whose value is uncertain
    • Exploitation: choosing the currently highest-valued action
    • Reward shaping: changing the objective after acting
    • Imitation: copying a labeled action
  2. 奖励分为三步，分别为 2、0 和 5。折扣系数为 0.9，第一步的回报是多少？
    • 7
    • 6.05
    • 6.5
    • 4.05
  3. 设计师提高了折扣回报目标的伽玛值。有什么变化？
    • The agent stops exploring
    • The reward function becomes a probability distribution
    • Future rewards count more relative to immediate rewards
    • Past rewards are counted again
  4. 策略只有在一系列操作之后才会获得奖励。出现哪个学分分配问题？
    • Which input column has the longest name?
    • Whether every action must receive the same reward
    • Whether the final reward erases all prior observations
    • Which earlier actions contributed to the eventual outcome
  5. 游戏代理通过收集无害的奖金来赚取积分，然后忽略任务，因为奖金很容易重复。这说明了什么失败？
    • Reward hacking: optimizing a proxy reward in a way that misses the intended goal
    • 灾难性遗忘
    • Exploration collapse caused by gamma
    • A transition model error
  6. A robot policy performs well in simulation. Why does it still need evaluation for its intended real deployment?
    • Simulation success proves every real-world action is safe
    • The simulator may rely on assumptions that fail outside it
    • A simulation prevents all reward loopholes
    • Real observations always match simulated ones
  7. A delivery task has delayed payoffs: a longer route now may prevent a costly delay later. Why does a one-step reward view miss the trade-off?
    • It assumes every action is random
    • It cannot represent observations
    • It ignores consequences that arrive in later states
    • It forces rewards to be positive
  8. A team changes the reward from “deliver safely” to “deliver quickly” and sees faster but riskier behavior. What should it conclude?
    • A faster policy proves the original safety objective was met
    • The discount factor alone guarantees safe behavior
    • Human preferences are unnecessary once training converges
    • Reward design affects behavior, so safety must be represented and evaluated explicitly