Backpropagation neural network with new improved error function and activation function for classification problem

Neural network has been used extensively for classification and many real world applications. The most commonly used neural network is multilayer perceptron with backpropagation (BP) algorithm. However the major problem of this algorithm is slow convergence rate and trap to local minima. The convergence is dependent on network parameters such as learning rate, momentum term and slope of activation function as well as its error function. This study proposes a New Improved BP algorithm which applies adaptive activation function using arctangent function in input-to-hidden layer and sigmoid logistic function in hidden-to-output layer. The efficiency and accuracy of the new improved method have been implemented and tested on two benchmark datasets: XOR and Balloon. The results show that the proposed method improved the convergence speed. However the classification accuracy is not very encouraging.