Constrained domain maximum likelihood estimation and the loss function in statistical pattern recognition

Abstract

In this thesis we present a new estimation algorithm for statistical models which does not incurs in the over-trainning problems. This new estimation techinque, the so-called, constrained domain maximum likelihood estimation (CDMLE) holds all the theoretical properties of the maximum likelihood estimation and furthermore it does not provides overtrained parameter sets. On the other hand, the impliations of the the 0-1 loss function assumption are analysed in the pattern recognition tasks. Specifically, more versatile functions are designed without increasing the optimal classification rule costs. This approach is applied to the statistical machine translation problem.