Many tuning methods are based on concepts of sensitivity analysis combined with heuristics that tend to reduce the search space by eliminating less promising configurations. Nevertheless, tuning parameters is a task that requires specific and timeconsuming experiments, especially when involving large problem instances. This is particularly due to existing methods were not designed to efficiently generalise a tuning of parameters to other instances that did not participate of the training process. In this paper, the recently proposed tuning method named Cross-Validated Racing (CVR) is revised in order to clarify theoretical fundamentals of tuning problem and expand the experiments to make possible evaluating its generalisation capacity against the reduction in the size of the training set. For validation, the Biased Random-Key Evolutionary Clustering Search (BRKeCS) is applied to solve scalable instance groups of Permutation Flow Shop Scheduling Problem. The computation results have demonstrated that CVR is robust in finding an effective parameter setting, requiring training process in only a half of total instance set.

No abstract

No abstract