LEARNING METHOD FOR NEURAL NETWORK, AND SALES PREDICTING DEVICE

摘要

PURPOSE: To prevent a large prediction error from being generated by making the neural network learn so that the total of specific square errors becomes minimum when future time-series data are predicted by inputting the actual result values of time-series data to the neural network. CONSTITUTION: By the learning method of the prediction device which predicts future time-series data by inputting the actual result values of time-series data to the neural network, the neural network is made to learn so that the total of square errors En no represented by the equations becomes minimum. In the equations, Yn is an actual result value at time (n), On-1 a predicted value at time n-1, On a predicted value at the time (n), On+1 a predicted value at the time n+1, and (k) a smoothing coefficient (provided that ok1). Consequently, when the article replenishment period prediction device of, for example, an automatic vending machine predicts the total number of sold articles, etc., by using the neural network, a large prediction error is prevented from being generated.