In the application field of production, scheduling with dispatching rules is facing the problem that no rule performs globally better than any other. Therefore, machine learning techniques can be used to calculate estimates of rule performances and select the best rule for each system state. A number of estimates are of poor quality and lead to a wrong selection of rules. Motivated by this problem, to further stabilize the selection approach a general approach, to automatically detect ‘faulty’ estimates from regression models is introduced and analyzed in this paper. Therefore, different models are learned and if their estimates differ strongly, it is likely that at least one model delivers poor estimates.
Additionally, a difference-threshold for our example data is defined. As a machine learning technique, we use Gaussian process regression with different covariance functions (kernels). The results have shown that our automatic detection