Being the study device which does the approximation of time series forecasting function

摘要

PPROBLEM TO BE SOLVED: To learn a prediction function of time series data on a non-Markov process by a continuous-valued function approximation technique ensuring short time convergence to a global solution. PSOLUTION: For the prediction of a time series of a non-Markov process, a function F handles an (n+m)-dimensional state äzSBt/SB} where m-dimensional context information äcSBt/SB} is added to an n-dimensional learning sample äxSBt/SB}, as time series information. Since the function F is the object of learning and the context information äcSBt/SB} is unknown, the estimation of the context information äcSBt/SB} and the learning of the function F are repeated alternately to provide asymptotic approach to an ideal solution. A learning sample is predicted with the learned function F, and a learning end determination is made with an error between the predicted value and an actually input learning sample. PCOPYRIGHT: (C)2007,JPO&INPIT