Model selection is the process of choosing classifier parameters suitable for the classification task at hand. In most cases this is done manually in an experimental fashion.
The search for the optimal model can also be automated. In Praat this is done by means of the same greedy local search algorithm used to search the weight space for feature weights. The model selection search implementation in Praat lets the user limit the search space with respect to the parameter k. By setting a maximum allowed value of k the search space can be shrunk considerably.
Due to its discrete (k) and nominal (vote weighting) nature, the size of the search space is normally of no concern, making an experimental/manual search tractable. The model selection feature of Praat becomes an essential tool only when applied to huge instance bases where the expected optimal value of k is high. In most cases however, manual experimenting will suffice.
© Ola Söder, May 29, 2008