METHOD AND DEVICE FOR GLOBAL ACCELERATION LEARNING FOR NEURAL CIRCUIT NETWORK MODEL

摘要

PROBLEM TO BE SOLVED: To shorten the search time by repeatedly executing a global search when a solution having an evaluation function value smaller than the evaluation function value of local solution can not be searched. ;SOLUTION: A correction amount calculation part 207 calculates the correction amount of weight from the difference of weight or gradient transferred according to the command of correction judge part 206, a weight correction part 208 corrects the value of weight, and an extremal value search convergence judging part 209 judges whether the stop preference of extremal value search is satisfied or not. A global search part 210 executes minimization to an evaluation function and searches a point where the evaluation function value gets smaller. A global search convergence judging part 211 judges whether the point found as a result of global search satisfies the stop recurrence of macro optimization on not. When the point of smaller evaluation function value is found out, a limit condition coefficient change part 212 changes the coefficient of limit condition. When the point of smaller evaluation function value is not found out, a global optimization coefficient change part 213 updates the coefficient of global optimization and when each processing is not finished even after repetition, processing is finished.;COPYRIGHT: (C)1998,JPO