概要:

Chaotic neuron has the temporally synthesizing ability. This study took time-series data and made their predictions by taking advantage of this ability. For learning, back-propagation was used.

The back-propagation and chaotic neuron have several parameters. But these parameters don't have proper values. So these parameters was determined by using genetic algorithms.

Time-series data used for learning are a sine wave and square waves. The results of the experiments showed that some predictions succeeded and the other failed.

It was investigated whether the number of the neurons which a neural network's middle layer has would influence the success rate of prediction. The results of the experiments showed that it does not affected the success rate of prediction so much.

The following nature was confirmed through the experiments. It's difficult to learn when time-series data used for learning, which have a certain length, include small changes and rapid changes. Whether or not to succeed in prediction is determined not only by the error near the end of the back-propagation. The appropriate parameters vary according to the form of time-series data used for learning. Neural network's middle class's structure does not affected learning so much.

Therefore, to search parameters by using genetic algorithm is effective. However, learning by changing the middle layer's number of neurons is not effective.