Co-evolutionary automatically defined functions in genetic programming

We show how the addition of co-evolution to genetic programming (GP) overcomes the current limitations of GP as well as GP augmented with automatically defined functions (GP+ADF) with a method called co-evolutionary automatically defined functions (GP+CADF). We demonstrate that GP+CADF requires a lower computational effort to solve the parity, sum of bits, image recognition, lawn coverage and the bumblebee problems. To further improve GP+CADF, we discover that using elitism lowers the computational effort required. We also discover ways to improve the initial population and initial best individuals used for evaluation.