A global optimization algorithm based on novel interval analysis for training neural networks

A global optimal algorithm based on novel interval analysis was proposed for Feedforward neural networks (FNN). When FNN are trained with BP algorithm, there exists some local minimal points in error function, which make FNN training failed. In that case, interval analysis was took into FNN to work out the global minimal point. For interval FNN algorithm, an interval extension model was presented, which creates a narrower interval domain. And more, in the FNN training, hybrid strategy was employed in discard methods to accelerate the algorithm's convergence. In the proposed algorithm, the objective function gradient was utilized sufficiently to reduce the training time in both interval extension and discard methods procedure. At last, simulation experiments show the new interval FNN algorithm's availability.