Abstract
A coordinate direction search algorithm is designed to train artificial neural network error function. The algorithm searches all possible directions in the error space. An acceleration step is introduced for quick convergence. The step is taken when successive search by the algorithm reduces the function value. The repeated successful search directions provide information for orthogonal move. This direction of search is defined as leap-frog step. The algorithm is suitable when complex geometry of the error surface is present in the form of stiff ridges, valleys, contours, or flat surfaces. Quite often derivative-based training algorithm terminates in local minimum. The leaf-frog step allows the algorithm to escape local minimum. The algorithm is derivative free and is convenient when the derivative information of an error function is not available. The algorithm converges to minimum value and is robust. This algorithm is a different class and is not a random search or a heuristic optimization method. It is quite different from the first- and second- order derivative-based training methods. The algorithm finds optimized neural network weights. It is tested with seasonal time series and classification problems.