A new algorithm is proposed to enable local minima to be escaped from by ch
anging the weights of a multi-layer neural network in a deterministic way.
Simulation results show that the new algorithm always outperforms other tra
ditional methods in achieving global convergence.