An improved algorithm of Neural Networks with Cubic Spline Weight Function

The paper proposes an improved Neural Networks construction with Cubic Spline Weight Function and its algorithm for the characteristic of the poor extending ability of the Neural Networks with Cubic Spline Weight Function. The Weight Function is divided into two parts by the improved algorithm. The Weight Function is trained by the three Cubic Spline Weight Function of the original algorithm and the constant coefficient of the Weight Function is trained by the grad dropping method. Because the new algorithm combines the merits of the Cubic Spline Weight Function Neural Networks with the merits of the traditional Neural Networks, the problems of the traditional dropping algorithm, such as the local minimum, slow convergence rate and initial value sensitivity, are not existed and the extending ability is better. The results of the simulation shows that compared to the traditional algorithm, the algorithm has high precision, fast speed and the extending ability remarkably improved compared to the unimproved algorithm.