On the computation of a robusl version of the optimal c(α) test