To provide the preparing method of model structure which can perform state division automatically and can stop the state division automatically.
The preparing device of model structure includes a storage device which stores a Hidden Markov Model (HMM), division parts S114, S120 which prepare respectively a new HMM by diving a certain state contained in the HMM in the direction of time and in the direction of environment, division selecting parts S116, S118, S122 to 126 which compute respectively the difference of a value of a prescribed evaluation function which is to be calculated by a Variational Bayesian method with respect to HMMs before and after the division and update the HMM by selecting the division in which the difference satisfies a prescribed selection condition and a control part S128 which makes S114 to S126, S130 repeatedly execute so as to repeat division until prescribed completion criteria are satisfied.
COPYRIGHT: (C)2005,JPO&NCIPI
Tetsu Nakamura
渡部 晋治、南 泰浩、中村 篤、上田 修功,ベイズ的基準を用いた状態共有型HMM構造の選択,電子情報通信学会論文誌,日本,社団法人電子情報通信学会,2003年 6月 1日,Vol.J86-D-II No.6,pp.776-786