Modeling global temperature changes with genetic programming

We use genetic programming (GP), a variant of evolutionary computation, to build interpretable models of global mean temperature as a function of natural and anthropogenic forcings. In contrast to the conventional approach, which engages models that are physically-based but very data-demanding and computation-intense, the proposed method is a data-driven randomized search algorithm capable of inducing a model from moderate amount of training data at reasonable computational cost. GP maintains a population of models and recombines them iteratively to improve their performance meant as an ability to explain the training data. Each model is a multiple input–single output arithmetic expression built of a predefined set of elementary components. Inputs include external climate forcings, such as solar activity, volcanic eruptions, composition of the atmosphere (greenhouse gas concentration and aerosols), and indices of internal variability (oscillations in the Ocean-Atmosphere system), while the output is the large-scale temperature. We used the data from the period 1900–1999 for training and the period 2000–2009 for testing, and employed two quality measures: mean absolute error and correlation coefficient. The experiment showed that the models evolved by GP are capable to predict, based exclusively on non-temperature data, the global temperature more accurately than a reference approach known in the literature.