Learning in Stages: A Layered Learning Approach for Genetic Programming

In this paper, we propose a new layered learning approach for Genetic Programming (GP), called GPLL. Our new GPLL is an extension of the earlier work in [8] incorporating theoretically and experimentally founded components derived from progressive sampling (PS). This new version of GPLL is tested and compared with standard GP on three real-world problems. Tuned for computational efficiency, it is able to demonstrate very substantial reductions in computational cost for relatively small (and generally non-significant) reductions in generalisation accuracy. At the other extreme, computational costs are still substantially less than for GP, while generalisation accuracies are consistently slightly better.