Research And Application Of Neural Network Classifier Based On The Soft Boundary

Classification is a significant problem in machine learning and statistics.In supervised classification tasks,the classifier learns knowledge from training samples,and tunes own action to obtain better classification performance.Scientists propose lots of methods to solve the classification problem,such as support vector machine,neural network,naive bayes and decision tree.Among these methods,the neural network has reached substantial success in realworld problems due to its universal approximation ability for nonlinear continuous functions.The classification process of a neural network can be described from the geometric interpretation.In the classification,the unknown sample is mapped from original data space to the partition space using the neural network.In the partition space,each class is presented by a point called centroid.The mapped sample is assigned to the closest centroid(class).However,the fixed centroid problem,which refers to that the position,number and label of the centroid are priori set,restrains the range of the optimal neural network.A neural network is success,if it can map the sample to correct class,even though to a class which centroid is located in an arbitrary position.The floating centroid method has been proposed to tackle the fixed centroid problem.In the FCM,the centroids are generated in accordance with the distribution of samples.Nevertheless,the hard boundary also restrains the neural network achieving better performance for the classification problem.The hard boundary problem refers to that the decision boundary is clearly described,and it results in a point,such as noisy or boundary,to be assigned to a class exclusively.The misclassification of noisy or boundary points causes the optimization fluctuate in the training of the neural network,and reduces the probability of finding the optimal neural network.In order to solve the hard boundary problem,the fuzzy logical theory is employed to assist the floating centroid method to generate the flexibility boundary.The mainly works of this study are described as follows.(1)The fuzzy floating centroid method(FFCM),which can generate the flexibility boundary,has been proposed to tackle the hard boundary.The FFCM combines the concept of floating centroid and fuzzy strategy to generate the flexibility boundary, which increases the probability of finding the optimal neural network in training process.Moreover,a class weight strategy has been proposed to assist the FFCM improve the performance for imbalance problem.(2)The fuzzy floating centroid method is employed to classify the cement strength grade.The measured strength grade varies among different specimens with the same configuration or even different regions of the same specimen.This phenomenon results in the disturbance of measured grade and increases the number of noisy and boundary points.FFCM has robust for the noisy and boundary points.Therefore,it is used to classify the cement strength grade.(3)The dynamic multilayer particle swarm optimization(DMLPSO)method has been proposed to optimize the floating centroid method.The optimization of neural network can be viewed as a complex multimodal function optimization problem.The DMLPSO has the ability to search the multimodal region thoroughly by using the multi-layer searching strategy and dynamic reorganizing strategy.Therefore,the DMLPSO can assist the FCM to obtain the neural network with the expected mapping performance.