Mathematical considerations and improvement on the fuzzy vector quantization‐based hidden markov model
Eiichi Tsuboka, Jun’ichi Nakahashi · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 1995
Abstract Two new types of fuzzy vector quantizationbased/hidden Markov models (FVQ/HMM)—multiplication‐type and addition‐type—are formulated to remove a mathematical inconsistency in conventional FVQ/HMMs proposed by Tseng et al. Experiments show the MTFVQ/HMM gives the best recognition rate among VQ‐type HMMs. Lettingytbe an observation vector at timet, C1, …,CMbe clusters to whichytis classified andsibe thei‐the state of HMM, we show that MT‐FVQ/HMM is derivable by defining the occurrence degree ofytatsito be the negative of the Kullback‐Leibler divergence ofa posterioriprobability distribution {P(C1|yt), …P(CM|yt)} froma prioriprobability distribution {P(C1|si), …,P(CM|Si)}.