Effect of parameter value and initial weights on the performance of backpropagation algorithm

Summary form only given, as follows. A feedforward neural network learning XOR task was used to demonstrate the influence of initial weights on the convergence behavior of the algorithm. Using selected initial weight sets, a series of deterministic experiments was then performed to determine the effect of learning rate and momentum on the convergence of the algorithm and time of convergence. The results show that there is a trade-off in choosing parameter values for faster convergence and that multiple global minima exist in the error surface