Evolutionary Selection of Regressional Predictors to Enhance the Performance of Microfossil-Based Paleotemperture Proxies

Using microfossil-based transfer functions, domain scientists from the field of pale oceanography seek to reconstruct environmental conditions at various times in the past. This is accomplished by first determining a quantitative relationship between a forcing function, such as temperature, and the modern for aminiferal response using a calibration data set based on environmental data from an oceanographic atlas and faunas generally extracted from sediment core tops. The method can be employed with a variety of environmental variables, but reconstruction of surface temperature is often the objective. The relationship developed using this training or calibration data set is then applied to down core data to infer past environmental conditions. The statistical methods that have been previously applied in this area can be grouped into three categories: linear regression based approaches, locally weighted regressions and neural networks. In addition to introducing some other learning algorithms including regression trees, bagging trees, random forest and support vector regression to this domain, in this study we suggest the use of model combination approaches to enhance the precision of estimation. By initializing with a pool of diverse predictors using a variety of learning algorithms and different samplings from the training and attribute set, a genetic algorithm was applied to select the best team of predictors. The optimal team was dominated by artificial neural network predictors suggesting their superiority over other methods tested with this type of data. The results also show the efficacy of the proposed approach compared to the other models.