Extension of Decoding Problem of HMM Based on LP-Norm
Abstract
The decoding problem of hidden Markov model (HMM) is extended based on the L <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</sup> -norm of a vector of the log transition probabilities along the sequence of hidden states. The extended decoding problem coincides with the conventional decoding problem for p = 1, and with the minimax decoding problem for p =∞. To solve the extended decoding problem, we introduce a family of Viterbi algorithm termed the “L <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</sup> -Viterbi algorithm” that continuously interpolates the conventional Viterbi algorithm and the minimax Viterbi algorithm. We also consider the corresponding evaluation and estimation problems. Numerical simulations show that the L <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</sup> -Viterbi algorithm with an adequately large value of p has an advantage over the minimax Viterbi algorithm.
BibTeX
@inproceedings{icassp2018_extensionofdecod,
title = {Extension of Decoding Problem of HMM Based on LP-Norm},
author = {Gen Hori},
booktitle = {ICASSP 2018},
year = {2018}
}