| 摘要 |
PROBLEM TO BE SOLVED: To provide a preprocessor in neural network learning for achieving the end determination of learning by appropriately recognizing total error evaluation functions even in the case of contradictory teacher data, and for acquiring necessary calculating precision in the case of ending learning. SOLUTION: The preprocessor is provided with: a pre-processing part for performing pre-processing to search the same input pattern in teacher data, and to count the number of cases of each input pattern, to calculate the mean value of the teacher output data of whole output patterns with respect to the same input pattern in the teacher data, to rewrite the mean value to the teacher output value of each input pattern, and to delete the overlapped section of the same input pattern data from the teacher input data; an error correction part for executing neural network learning on the basis of teacher data processed by the pre-processing part, and for deriving the error evaluation functions of each input pattern; and a total error evaluation function part for calculating a total error evaluation function for determining the convergence conditions of the neural network learning by using the error evaluation function of each input pattern to be derived from the error correction part and a weighting factor configured of the number of cases of each input pattern. COPYRIGHT: (C)2011,JPO&INPIT
|
