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。
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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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