Variations in Optimality Theory: Simulated Annealing and other methods. Conference talk given at: TABU-day, June 3, 2005, Groningen, Netherlands. Slides. With Dicky Gilbers and Maartje Schreuder.
Traditionally, Optimality Theory (OT) accounts for language typology: each constraint ranking corresponds to a different language type, by making a different candidate the optimal element of the candidate set. Consequently, one grammar returns one output. Languages, however, include variation (free alternation, register-dependent alternation, fast speech errors, etc.). How could one account for these phenomena within OT?
The present talk introduces "Simulated Annealing Optimality Theory" (SA-OT), an algorithm searching for the optimal element of the candidate set. Yet, the search does not always return the desired element, yielding a "performance error". Subsequently, SA-OT is compared to other mutations of OT allowing for more than one outputs, such as the proposal of Anttila and Cho (1998), Stochastic Optimality Theory (Boersma and Hayes, 1999) and the proposal of Coetzee (2004). After the comparison of the predictions of
the different models, we shall argue for the adequateness of SA-OT, based on empirical data by Maartje Schreuder.