Genetic Programming–A Preliminary Study of Knowledge Transfer in Mutation

Abstract

In this research, we introduce two new variations of Genetic Programming (GP) that reusing transferred knowledge in the mutation operator. They are called FullTree.CM_1 and FullTree.CM_2. Chromosomes of the best individuals learned from previous source problems will be reused as sub-trees of children generated by mutation while training GP for the target problem. FullTree.CM_1 takes k% good individuals in the last generation of a previous learned sub-task (source problem) into a POOL for reusing in the mutating process while FullTree.CM_2 re-uses k% good solutions from all previous learned sub-tasks. In order to analyzing the effectiveness of these schemes, we use three evaluation criteria including errors on training data, errors on testing data, and size of the learned solutions. Experimental results show that our schemes perform better than other previous transfer GP on the most of tested problem families with all of these three evaluation criteria. This is very promising and motivating for further research to improve GP based on transfer mutation.