A New Technique for Searching the Global Minimum of Supervised Neural Network

This paper presents a technique in how to searching the global minimum for the supervised neural network training. This technique is developed based on the idea of nearly equivalent model. To demonstrate the new technique proposed, two signal processing studies, including signal recognition and signal modeling were simulated. For a comparison, the same simulations were also performed by using the neural network with the standard steepest descent error back-propagation (BP) algorithm. From the simulation results shown, the technique we proposed not only can evidence whether the neural network is in the local training or not, but also can show that the “best” performance of the neural network should have.