| 摘要 |
PURPOSE: To conduct enhanced leaning by combining a short-time operation by the (+) net and a short-time operation by the (-) net as one cycle, repeating the operations of prescribed cycles, comparing the allocation performance of respective neural networks based on the statistical result of the prescribed cycles, and updating the bonding weight. CONSTITUTION: A precession is applied to the bonding weight of an allocating neural network (S2), and multiple nets having different precession quantities are prepared. The (+) net applied with a precession to the (+) side and the (-) net applied with a precession to the (-) side are prepared respectively (S3), (S4), for example. A short-time operation allocated to the (+) net and a short-time operation allocated to the (-) neural net are combined as one cycle, and operations of prescribed cycles are repeated. The allocation performance by the (+) net and the allocation performance by the (-) net during that period are compared (S13). The precession of the net having the superior performance is applied to the bonding weight of the allocating neural network for updating. |
