GLOBAL OPTIMIZATION IN NEURAL-NETWORK TRAINING

This paper presents a nerv neural network training scheme for pattern recognition applications. Our training technique is a hybrid scheme wh ich involves, firstly, the use of the efficient BFGS optimisation meth od for locating minima of the total error function and, secondly the u se of genetic algorithms for finding a global minimum. This paper also describes experiments that compare the performance of our scheme with three other hybrid schemes of this kind when applied to challenging p attern recognition problems. Experiments have shown that our scheme gi ves better results than others.