TRIAM-1Mにおけるフィードバック制御に用いるニューラルネットワークの構造最適化

概要

論文の詳細を見る
A neural network for feedback control must be compact for insuring fast executing with high accuracy. A procedure of determining structures of input signals and 2 hidden layers is developed in order to construct the network for TRIAM-1M feedback control with high accuracy and high speed. From relation between degradation of neural network performance and effect to output by removing unit, two expressions of redundancy is necessary. The structures depend on an expression of redundancy and a process of the determination. The most useful procedure is iteration of removing the most redundant unit until the redundancy becomes under a criterion of removal. The number of units in the network developed can be reduced to about 60% of units used in the original network without the degradation of the accuracy. The validity of the procedure is verified by the neural networks, which is constructed to calculate plasma position and learned with the supervisor data is calculated by equilibrium code on TRIAM-1M.
社団法人プラズマ・核融合学会の論文
2000-09-25