概要:

The hierarchical neural network with internal memory layer is suitable for learning of time-series. The conventional back-propagetion was not effective for learning the network. The newly devised ``delayed learning method'' enables the neural network with internal memory to learn complicated time-series.

This study took temperature fluctuations as an example of a complicated time-series for the network to learn. The influence on the temperature forecast by the scale of the delay time and the network is examined, and the decision method of effective parameters for learning a complicated time-series was considered and discussed. As the result, the learning succeeded under the condition of the large network size and the short delay time, regardless of the supervisory signal.