引導式測驗 · 中等 等級

超參數測驗的貝葉斯優化

當貝葉斯優化選擇昂貴的模型試驗時,了解替代模型、取得權衡和洩漏安全驗證。

相關引導路徑超參數的貝葉斯優化
問題 1 的 8

代理模型在貝氏最佳化中代表什麼?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. 代理模型在貝氏最佳化中代表什麼?
    • A probabilistic approximation of performance across hyperparameter settings
    • The final trained model's input data
    • A fixed grid of every configuration
    • A causal diagram of feature effects
  2. 擷取函數有助於選擇什麼?
    • Which examples to delete from the test set
    • Which hyperparameter configuration to evaluate next
    • The final model's class labels
    • The number of training observations
  3. 預期改善如何平衡探索與開發?
    • It selects the most uncertain point regardless of performance
    • It evaluates only the setting with the best predicted mean
    • It considers expected gain over the current best, including uncertainty
    • It maximizes the training set size
  4. 為什麼在超參數搜尋期間要保留最終測試集?
    • The final test score sets the acquisition function
    • The test set improves surrogate smoothness automatically
    • Bayesian optimization cannot use validation data
    • Repeated tuning on test results leaks information into model selection
  5. 選擇不當的搜尋空間會造成什麼影響?
    • Guarantee the optimum is found
    • Exclude strong configurations or waste trials on irrelevant regions
    • Remove evaluation noise
    • Make validation unnecessary
  6. Why repeat noisy optimization with different seeds?
    • It changes categorical settings into continuous ones
    • It makes all trials independent of validation scores
    • The trial path and training noise can change the selected setting
    • It guarantees the best possible score
  7. Which role can a Gaussian process surrogate provide?
    • A replacement for the held-out test set
    • Exact performance for every untried configuration
    • A rule that prevents overfitting by itself
    • Predictive mean and uncertainty for candidate settings
  8. What should a search report include for reproducibility?
    • Bounds, trial budget, validation design and optimization settings
    • Only the single best score
    • Every test score used during tuning
    • The alphabetical list of parameter names