Novel Application of Mutual Information in Transfer Learning for Genetic Programming
Yilin Liu, Gareth Anthony Taylor, Zhengwen Huang · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2025
Genetic Programming (GP) faces two major challenges: the inability to use knowledge from past problems, as each run is independent, and the difficulty in identifying building blocks in the early stages of evolution. These limitations result in high computational costs and slower convergence. Transfer learning can provide a solution by leveraging past experiences to enhance performance in GP. A mutual-information-based transfer learning method is proposed in this paper to identify and transfer beneficial knowledge fragments. The method is evaluated on ten polynomial and trigonometric symbolic regression problems from previous literature and compared with standard GP and SubTree50 as state-of-the-art methods. Results of the above experiment demonstrate improved or comparable performance of the proposed method in terms of accuracy, statistical significance, and generalization. The contribution of this paper includes: a mutual-information-based technique for identifying and transferring knowledge fragments. Results highlight the potential of mutual information to address key challenges in GP effectively.