| 摘要 |
PURPOSE: A multilayered neural network learning method is provided to reduce a convergence time of an error back-propagation learning algorithm by using a continuous weight modification routine. CONSTITUTION: A multilayered neural network learning method comprises steps of initializing a weight value and an allowance error, and setting a WMRV(Weight Modification Repetition Variable)(S2), inputting a learning pattern and calculating responses from each layer(S4), calculating a response from an output layer and an error from a hidden layer(S6,8), adjusting the weight value between the hidden layer and the output layer by the WMRV(S10), determining if a learning pattern exists and comparing an error square with a maximum error square in the case that the learning pattern does not exist(S12,14), returning to the step S4 in the case that the learning pattern exists, returning to the step S4 if the error square is not less than the maximum error square(S16), and stopping the process if the error square is less than the maximum error square.
|
