Dynamic tunneling technique for efficient training of multilayer perceptrons

A new efficient computational technique for training of multilayer feedforw ard neural networks is proposed. The proposed algorithm consists two learni ng phases. The first phase is a local search which implements gradient desc ent, and the second phase is a direct search scheme which implements dynami c tunneling in weight space avoiding the local trap thereby generates the p oint of next descent. The repeated application of these two phases alternat ely forms a new training procedure which results into a global minimum poin t from any arbitrary initial choice in the weight space. The simulation res ults are provided for five test examples to demonstrate the efficiency of t he proposed method which overcomes the problem of initialization and local minimum point in multilayer perceptrons.