Hybrid Transfer Active Learning for Multistream Processes With Within-Process and Cross-Process Correlation Modeling...

Area of Science:

• Machine Learning
• Statistical Modeling

Background:

• Active learning for regression (ALR) is crucial for learning functional relationships.
• Existing ALR methods face challenges with the cold-start problem and single-stream limitations.

Purpose of the Study:

• To propose a hybrid transfer learning framework for multistream ALR.
• To address the cold-start problem by modeling within-process and cross-process functional correlations.

Main Methods:

• Introduced a novel multioutput Gaussian process (MGP) covariance structure.
• Integrated offline learning (cross-process knowledge transfer) and online updating (within-process refinement).

Main Results:

• Demonstrated monotonic decrease in ALR error with data acquisition.
• Showcased the framework's accuracy and superiority over benchmark methods.

Conclusions:

• The proposed framework effectively resolves the cold-start problem in multistream ALR.
• Hybrid transfer learning with MGP provides a robust approach for complex functional learning.