with the least MSE reduction, that is, perlbmk, the MSE reduction can still achieve 35%. Comparing with ANN-IS, COAL reduces the MSE by 51% in average. Hence, we can conclude that COAL is much more practical than state-of-the-art DSE techniques due to its high accuracy.
Figure 3 also presents the comparison between supervised M5P regression tree and COAL. It can be observed that COAL reduces the MSEs of M5P over all 14 benchmarks. In average, COAL reduces the MSE by 42% of M5P. Hence, it can be concluded that the exploitation of unlabeled configurations in COAL can significantly improve the prediction accuracy.