A New Early Stopping Algorithm for Improving Neural Network Generalization

As generalization ability of neural network was restricted by overfitting problem in the network´s training. Early stopping algorithm based on fuzzy clustering was put forward to solve this problem in this paper. Subtractive clustering and fuzzy c-means clustering (FCM) were combined to realize optimal division of training set, validation set and test set. How to realize this algorithm in backpropagation (BP) network by utilizing neural network toolbox and fuzzy logic toolbox in MATLAB was dwelled on. Early stopping algorithm based on fuzzy clustering and other early stopping algorithms were applied in function approximation and pattern recognition problems in validation experiments. Experiments results indicate that early stopping algorithm based on fuzzy clustering has higher precision in comparison to other early stopping algorithms. Outputs of training set, validation set and test set are more accordant.