An algorithm to determine neural network hidden layer size and weight coefficients

The strictly decreasing relationship between the sample approximation error and the number of hidden units in a three layer artificial feedforward neural network (AFNN) is proven in the sample space. The relationship is a powerful tool in determining the number of hidden units needed. A hybrid optimization algorithm is proposed on the relationship for simultaneously determining the number of hidden units and weight coefficients in the AFNN. The algorithm is the synthesis of golden section, evolutionary programming and gradient based algorithm which is effective in determining the number of hidden units and weight coefficients in the neural network.