Further investigation on genetic programming with transfer learning for symbolic regression

Abstract

Transfer learning is an important approach in machine learning, which aims to solve a problem by utilising the knowledge learnt from another problem domain. There has been extensive research and great achievement on transfer learning for image analysis and other tasks, but research on transfer learning in genetic programming (GP) for symbolic regression is still in the very early stage. However, GP has a natural way of expressing knowledge by trees or subtrees, which can be automatically discovered during the evolutionary process. An initial work on GP with transfer learning was proposed to transfer knowledge through best trees or subtrees from to source domain to facilitate the learning in the target domain. However, there are still a number of important issues remaining not investigated. This paper further investigates the ability of GP with transfer learning on different types of transfer scenarios, investigates the influence of a key parameter and the effect of transfer learning on the evolutionary training process, and also analyses how the knowledge learnt from the source domain was utilised during the learning process on the target domain. The results show that GP with transfer learning can generally perform well on different types of transfer scenarios. The transferred knowledge can provide a good initial population for the GP learning on the target domain, speed up the convergence, and help obtain better final solutions. However, the benefits of transfer learning varies in different scenarios.