Genetic Programming for Instance Transfer Learning in Symbolic Regression.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Intelligence

Background:

• Transfer learning enhances model performance by leveraging knowledge from similar domains.
• Limited research exists on transfer learning within genetic programming (GP) for symbolic regression tasks.
• Existing methods often struggle with effective knowledge transfer and generalization.

Purpose of the Study:

• To propose a new instance weighting framework for transfer learning in genetic programming-based symbolic regression.
• To improve the identification and utilization of relevant source-domain data for enhanced learning.
• To effectively address the challenge of cross-domain generalization and overfitting in symbolic regression.

Main Methods:

• Developed a novel instance weighting framework for transfer learning in GP.
• Employed differential evolution to optimize weights for source-domain instances.
• Utilized density estimation for effective instance selection and to guide weight optimization.

Main Results:

• The proposed method significantly improved cross-domain generalization performance and stability.
• Effectively reduced the trend of overfitting compared to baseline methods.
• Evolved simpler symbolic regression models than traditional GP approaches.

Conclusions:

• The proposed instance weighting framework offers a robust solution for transfer learning in GP symbolic regression.
• This approach enhances model performance, generalization, and simplicity.
• It effectively addresses limitations in current transfer learning applications within GP.