गाइड से जुड़ी प्रश्नोत्तरी · मध्यम स्तर

लॉग्ड डेटा क्विज़ से ऑफ-पॉलिसी मूल्यांकन

व्युत्क्रम प्रवृत्ति भार लागू करें और ओवरलैप, विचरण और लॉगिंग आवश्यकताओं की पहचान करें जो प्रतितथ्यात्मक नीति अनुमानों को सीमित करते हैं।

संबंधित मार्गदर्शक पथऑफ पॉलिसी मूल्यांकन लॉग डेटा
प्रश्न 1 का 8

What does off-policy evaluation estimate?

प्रत्येक प्रश्न का पूर्वावलोकन करें8 प्रश्न

जैसे ही आप प्रश्नोत्तरी लेते हैं उत्तर और स्पष्टीकरण दिखाई देते हैं।

  1. What does off-policy evaluation estimate?
    • A target policy's outcomes using data generated by another policy
    • The training accuracy of the logging model
    • A model's causal feature effects
    • The total number of logged requests
  2. Why does IPS divide by the logging propensity?
    • To make rewards nonnegative
    • To reweight observed actions according to how likely the old policy was to choose them
    • To eliminate the need for outcomes
    • To estimate the context distribution only
  3. What happens when the logging policy assigns a very small probability to a chosen action?
    • The candidate policy is unbiased automatically
    • The action outcome becomes known with certainty
    • The inverse weight can become large and increase variance
    • The reward is set to zero
  4. What does lack of support mean for a target action?
    • Every reward is missing
    • The action is always optimal
    • The target policy equals the logger
    • The logs contain little or no evidence for that action in the context
  5. Why report effective sample size with weighted estimates?
    • It changes the propensity values
    • It reveals whether a few large weights dominate the information
    • It measures the number of policy actions possible
    • It proves the reward model is correct
  6. What tradeoff can clipping IPS weights create?
    • Guaranteed unbiasedness
    • Higher sample size with no assumptions
    • Lower variance but potential bias
    • No effect on the estimate
  7. What must be logged to support ordinary IPS evaluation?
    • The candidate policy's future decisions
    • Only model predictions
    • Only final reward totals
    • Action probabilities and outcomes with relevant context
  8. When might a controlled experiment be needed instead of relying on OPE?
    • When the candidate policy acts outside historical support
    • Whenever a metric is positive
    • When propensities are accurately logged
    • When the target matches the logging policy