In Optimality Theory (OT, Prince and Smolensky 1993/2004), the surface form is predicted to be the most harmonic (the optimal, the best) element of a candidate set that is generated from the underlying form. Hence, OT requires an optimization algorithm that is able to find the best element of the candidate set with respect to the target function called Harmony. In this talk, based on my dissertation ("Finding the Right Words", 2006, http://dissertations.ub.rug.nl/faculties/arts/2006/t.s.biro/, http://www.birot.hu/publications.php), I introduce the Simulated Annealing for Optimality Theory Algorithm (SA-OT), a variant of the well-known and wide-spread heuristic optimazation algorithm originating in statistical physics. The algorithm differs from standard simulated annealing because the target function is not real-valued; the consequences of this fact will be presented.
As the algorithm does not guarantee to always find the optimal candidate, the algorithm can be used as a model of linguistic performance. The errors made by the algorithm are interpreted as performance errors, for instance, as fast speech errors whose frequencies increase if the algorithm is run more quickly. The model is compared to Persistent OT ("harmonic serialism") by McCarthy (2006), as well as to the architecture proposed in The Harmonic Mind by Smolensky and Legendre (2006). We shall seriously question the latter's central claim, namely that "there is no symbolic algorithm whose internal structure can predict the time and the accuracy of processing; this can only be done with connectionist algorithms" (vol. 1, p. 91).