| 摘要 |
PURPOSE: A system for forecasting a typhoon using an evolution strategy and a bilinear regression neural net is provided to forecast a typhoon rapidly and accurately without using a super computer. CONSTITUTION: In a typhoon forecasting system having three neural nets for analyzing a latitude, a longitude, and a central atmospheric pressure for analyzing a central position and a central atmospheric pressure of a typhoon, plural weight values corresponded to three input data of a latitude, a longitude, and a central atmospheric pressure being inputted to each neural net are multiplied, and a blind layer is formed. The result of each blind layer is used as input components by applying a regression neural net. The optimum learning rate is decided through a repetitious experimentation for calculating a proper result in the regression neural net. Two activation function gradients of the regression neural net are calculated through an evolution strategy. If a learning error is increased during a progress, the learning is stopped and all weight values are calculated again by adjusting the gradients of the activation functions. The gradients of the activation functions are adjusted using an evolution strategy. Thus, a course and a central atmospheric pressure of an actual typhoon are forecasted from the inputted latitude, longitude, and central atmospheric pressure values using the optimum activation function gradient and each weight value.
|
