Differential Evolution for Instance based Transfer Learning in Genetic Programming for Symbolic Regression

Transfer learning attracts increasing attention in many fields in recent years. However, studies on transfer learning for symbolic regression are still rare. This work proposes a new instance weighting framework for genetic programming (GP) based symbolic regression for transfer learning. The key idea is to use differential evolution to search for optimal weights during the evolutionary process of GP, which helps GP identify and learn from more useful source domain instances and eliminate the effort of less useful source domain instances. The results show that the proposed method achieves notably better cross -domain generalisation performance in a very stable way than GP without the instance weighting framework and support vector regression.