Backpropagation neural networks are trained by adjusting initially random interconnecting weights according to the steepest local error surface gradient. The authors examine the practical implications of the arbitrary starting point on the error landscape of the ensuing trained network. The effects on network convergence and performance are tested empirically, varying parameters such as network size, training rate, transfer function and data representation. The data used are live process control data from an injection molding plant.
展开▼
