TRAINING METHOD FOR NEURAL NETWORK WITH TECHNIQUE-BASED NONDETERMINISTIC CHARACTERISTIC

摘要

PROBLEM TO BE SOLVED: To correctly adjust the weight of the neural network which should train characteristics of a technical system having high components of a probability event by evaluating whether or not a controlled variable is improved as to target characteristics of the technical system with a cost function and then promoting the weight adjustment with the cost function. SOLUTION: The neural network NNW controls the technical system (f). The neural network NNW sends an adjustment quantity Ut-1 out to the technical system (f) through a connection line 150. The technical system (f) generates a controlled variable according to this adjustment quantity Ut-1 . This controlled variable is supplied to a delay element Z1. Then a weight coefficient which is adjusted by the network is increased or decreased with the cost function as to advantageous system characteristics of the technical system (f). Thus, the adjustment quantity Ut-1 having random noise as to the statistical distribution of a time series is used to attain the weight adjustment state of the neural network NNW that the advantageous target characteristics of the technical system (f) operate on.