Optimal grid quantization

Optimal quantization, a non-parametric technique for pattern recognition, determines a compact and efficient density representation of data by optimizing a global quantizer performance measure, which is a weighted combination of average log likelihood, entropy and correct classification probability. In multidimensions, we obtain the quantization grid using genetic algorithms. Smoothing is an important aspect as it affects the generalization ability of the quantizer. We propose a fast k neighborhood smoothing algorithm. Optimal quantization is much more efficient than other non-parametric methods. For not very well separated Gaussian mixture models, it produces much better results than the EM algorithm, which fails to converge to the true parameters of the underlying density.