A new approach to global optimization and its applications to neural networks

A new approach to global optimization that alternately rocks the landscape of the objective function and rolls the ball representing the current state of the variable down to the bottom of the nearest valley is presented. The degree of perturbation is determined by a parameter called rock level. The rock level decreases in the process. The ball gets rocked out of local minima and eventually settles at a global minimum. Rock is affected by either perturbing the constants in the objective function or adding a perturbing function to it or both. Roll is performed by a local search. It is shown that the Hopfield net can be rocked to produce a combinatorially minimal solution and that the error backpropagation can be rocked to produce a globally optimal multilayer perceptron