抄録
The estimation of Markov jump processes from noisy observations is a nonlinear estimation problem with strong nonlinearity. A new approach to solve the estimation problem is presented, using a neural network model. The neural network is designed to minimize an energy function, which consists of two terms: one represents reliability of observation data and the other imposes penalties on estimates on the basis of a priori information about the processes to be estimated. It is shown that nearly optimal estimates are obtained using the neural network as a sliding window filter. The quality of the estimates depends on the ratio between two terms in the energy function. It is also shown that an adequate value of the ratio is learnable from samples of true processes and observation data using a stochastic approximation method.
