PARAMETER CONTROL METHOD AND GLOBAL OPTIMIZATION METHOD FOR NEURAL NETWORK

摘要

PURPOSE: To obtain an extremely satisfactory answer for an optimization problem at high speed by a definite method without using stochastic change in a neural network in which total input to each element and the state of each element are updated according to a specific equation by varying the value of a parameter within a specific range in a process to repeat such update. ;CONSTITUTION: In an intercoupling type neural network in which the output X_k of the element (k) is transmitted to the element (j) with weight W_jk, and W_jkX_k flows in the element (j), and also, on which input bias θ_j is added, and in which the total input U_j to each element and the state X_j(j=1,...n) of each element are updated according to equations I, II setting the parameter to control the influence of an output value before the state is updated as (r), and provided with a transfer function Sats like equation Ill assuming the parameter to control the inclination of the transfer function as (s), the values of the parameters (s) and (r) are varied within the range of equation IV in the process to repeat the state update of the element.;COPYRIGHT: (C)1994,JPO&Japio