RESULTS

In the California Bight data 14 samples were available. Half of these samples have been used as the training set (seven observations). The remaining half constituted the test set (also seven observations). To make sure that the error rate is representative of the dataset, the network was run five times, each with 50% random training-set and 50% test-set members (the observations for the training sets are in italics in Table 1). The number of network configurations attempted was 600 (20 generations of 30 populations each). Best network configurations for the various partitions are shown in Table 3. The average Root-Mean-Square-Error of Prediction (RMSEP) in the test sets was 0.68, implying that an unknown SST can be predicted from the relative abundances of the five nannoplankton species used here with a precision of ± 0.68°C (Figure 3). Correlation coefficients between observed and predicted SSTs in the five test sets range between 0.80 and 0.98 (Figure 4).

For the Mediterranean data the number of samples was 48. We trained the network using 80% of the samples as the training set (39 observations) and the remaining 20% as the test set (nine observations). This subdivision of the training and test sets was performed randomly by the program. Again, five different random subdivisions of the original data set were run to generate an error estimate. Also in this case, the number of network configurations attempted was 600 (20 generations of 30 populations each). The configurations of the best networks are provided in Table 3. The average RMSEP in the test sets was 0.64, implying that an unknown oxygen isotope value can be predicted from the relative abundances of the selected nannofossil species in this core with a precision of ± 0.64 ¹⁸O PDB (Figure 5). Correlation coefficients between observed and predicted isotope data in the five test-sets range between 0.64 and 0.96 with an average value of 0.88 (Figure 6).