The choice of training instances significantly impacts the performance of automated algorithm selection models for numerical black-box optimization problems.
