Naive Parameter Learning for Optimality Theory – the Hidden Structure Problem
Jarosz, Gaja. 2013. Naive Parameter Learning for Optimality Theory – the Hidden Structure Problem. In Proceedings of the 40th Annual Meeting of the North East Linguistic Society.
This paper introduces a new learning algorithm, NPRL, that outperforms RIP/GLA, RIP/SGA, and RIP/CD given the same learning conditions. Analysis of NPRL suggests that its successful performance can be attributed to an approximated random search. True random search outperforms all of the above algorithms, achieving perfect performance in fewer iterations on average than NPRL. This raises a number of questions regarding the evaluation and development of computational models of phonological learning.