Bi-directionalization of neural computing architecture for time series prediction. III. Application to laser intensity time record "Data Set A"

For part II, see Int. Conf. on Dynamical Aspects in Complex Systems from Cells to Brain, p.43-4 (2000). One of the most important targets of time series prediction is an improvement of prediction quality for aiming at prefect prediction. To reach the goal, most studies have used uni-directional computation flow to predict future events from present and past information. In this study, on the contrary, bi-directional computation style is applied to a time series prediction task to investigate its effectiveness. As a result of computer simulations with the laser intensity time record "Data Set A", it is clear that the coupling effect between the future and past prediction transformations produce a good advantage on trainability, generalization, and prediction quality over the conventional uni-directional network.