We fed short overt Latin stress patterns to 100 virtual language learners whose grammars consist of a universal set of 12 Optimality-Theoretic constraints. For 50 learners the learning algorithm was Error-Driven Constraint Demotion (EDCD), for the remaining 50 it was the Gradual Learning Algorithm (GLA). The EDCD learners did not succeed: they ended up in a grammar that could not reproduce the correct stress pattern. The GLA learners did succeed: they came up with an analysis close to one of the analyses proposed in the literature. These results add to previous findings that the GLA seems to be a more realistic ingredient than EDCD for models of actual language acquisition.