SYSTEM AND METHOD FOR TIME-SERIES TREND ESTIMATION BY RECURSIVE TYPE NEURAL NETWORK IN COLUMN STRUCTURE

摘要

PROBLEM TO BE SOLVED: To efficiently estimate the trend of time-series data which vary discontinuously by making the relation between the internal state of the neural network and the time-series data distinct. SOLUTION: The column structure recursive type neural network (CSSRNN) 19 is equipped with (m) columns consisting of neural elements 51-j (j=1,..., m) and (s) registers 52-j-k (k=1,..., S). Each neural element generates an output at time (t) from an input x(t) and each column passes the output history of the neural elements before the time (t) to a nonlinear equation solving device 18. The nonlinear equation solving device 18 finds the zero point of a target function from the passed history and calculates the probability density of the value (x) corresponding to each zero point. Then the value having the largest probability density is passed as a predicted value of input data at next time. The independency of each column is high and discontinuous discrete values are suitably predicted.