| 摘要 |
PROBLEM TO BE SOLVED: To produce a novel pattern which is not learnt. SOLUTION: Data x<SB>t</SB>corresponding to a prescribed time series pattern are inputted to an input layer 11 of a recurrent neural network 1, and a predictive value x*<SB>t+1</SB>is obtained from an output layer 13. A difference between data x<SB>t+1</SB>as a teacher and the predictive value x*<SB>t+1</SB>is learnt by a back propagation method, and a weighting coefficient of the intermediate layer 12 is set to a prescribed value. After a plurality of time series patterns are learnt, a parameter different from a value during learning is inputted to a parametric bias node 11-2, and a non-learnt time series pattern corresponding to the parameter is produced from the output layer 13. The invention is applied to a robot. COPYRIGHT: (C)2004,JPO
|
