Genetic Programming: Magic Bullet, Poisoned Chalice or Two-Headed Monster?

Is genetic programming (GP) a magic bullet, poisoned chalice or two-headed monster? This chapter unpacks these questions by equipping readers with the necessary information to both productively utilise this powerful data-driven modelling tool and to formulate some answers for themselves. Nevertheless, navigating the GP road map can be tricky, not least because at its heart, there are some weighty computational concepts that require suitable analogies and metaphors to make them more easily understood – for this, its inventors piggyback onto the field of biological evolution. The large number of borrowed terms and associated synonyms involved, however, should not in any way discourage the discerning scientist, since GP permits fresh and novel solutions to be evolved for resolving complex scientific problems. This chapter builds on an earlier account of GP presented in the first edition of this book by discussing recent major improvements to the accessibility of the technique, which is mainly driven by advances in computing technology and software innovation. The first section of the current chapter provides some background to GP and its key concepts. The central section provides a general framework for performing GP experiments, with symbolic regression being selected as the common thread from which we weave our story. This is followedAbstract ... 169 8.1 Key Facts on Genetic Programming ... 170 8.2 Symbolic Regression ... 173 8.3 Getting the Modelling Right... 1748.3.1 Stage 1 Study Site Selection ... 175 8.3.2 Stage 2 Data Preparation ... 175 8.3.3 Stage 3 Model Development ... 176 8.3.4 Stage 4 Rejecting and Accepting Models ... 181 8.3.5 Stage 5 Model Testing and Evaluation ... 1828.4 Case Studies ... 182 8.4.1 Estimating Pan Evaporation ... 182 8.4.2 Rainfall-Runoff Modelling ... 185 8.4.3 Building a Spatial Interaction Model... 1918.5 Future Directions ... 194 8.5.1 Is It Important How Things Are Modelled? ... 195 8.5.2 Should Models Be Complex? ... 196 8.5.3 Does It Matter What You Model? ... 196 8.5.4 Here Be Dragons? ... 197References ... 198by three simple case studies are presented with the intention of inspiring readers to develop and test their own ideas about what GP can be used for. Finally, a short discussion on the future directions of GP helps steer some searching philosophical questions for the reader to consider. One of the key messages from the first edition version of this chapter remains suitably apt – ‘if in doubt, experiment’! It is