主权项 |
1. A model estimation device comprising:
a data input unit which acquires observed value data subjected to estimation of a latent variable model, a type of each component constituting the latent variable model, and a state number candidate set of the latent variable model; a state number setting unit which sets an element for which computation has not been completed yet in the state number candidate set, as the number of states; an initialization unit which sets initial values of a variational probability of a latent variable, a parameter, the type of each component, and a free parameter selection variable; a latent variable variational probability computation unit which acquires the observed value data and the variational probability of the latent variable, the type of each component, the parameter, and the free parameter selection variable set by the initialization unit, and computes the variational probability of the latent variable so as to maximize a lower bound of a model posterior probability limited in degree of freedom by the free parameter selection variable; a component optimization unit which acquires the observed value data and the variational probability of the latent variable, the type of each component, the parameter, and the free parameter selection variable set by the initialization unit, and estimates an optimal type of each component and a parameter thereof by optimizing the type of each component and the parameter so as to maximize the lower bound of the model posterior probability limited in degree of freedom by the free parameter selection variable and separated for each component of the latent variable model; a free parameter selection variable computation unit which acquires the type of each component and the parameter thereof estimated by the component optimization unit and the variational probability of the latent variable computed by the latent variable variational probability computation unit, and computes the free parameter selection variable; an optimality determination unit which acquires the observed value data, the type of each component and the parameter thereof estimated by the component optimization unit, the variational probability of the latent variable computed by the latent variable variational probability computation unit, and the free parameter selection variable computed by the free parameter selection variable computation unit, and determines whether or not to continue the maximization of the lower bound of the model posterior probability limited in degree of freedom; and a result output unit which outputs the variational probability of the latent variable computed by the latent variable variational probability computation unit and the type of each component and the parameter thereof estimated by the component optimization unit. |