Optimization And Semi-supervised Learning Of The Covering Network

Traditional Neural Networks have some essential limitations such as intricate structure, poor performance and lower efficiency. The algorithm of SVM has great generalization ability but it is difficult to solve multi-class. Based on the geometrical significance of nerve cell Professor Zhang Ling etc proposed a constructive method for machine learning based on covering network. With such method, the structure of multi-class network is translated into finding a series of coverage areas in the sample set which can distinguish different class of points. The network of this method is easy to structure and can effectively deal with multi-class problem. Theoretical studies show that the performance of covering network is closely related with the number of coverage areas. When we structure coverage areas, the choice of the mid-points of coverage areas has a direct impact on the number of the areas. Meanwhile covering algorithm is a supervised learning algorithm, which requires a lot of labeled examples. It is very easy to obtain a large number of unlabeled examples, but for a large number of labeled ones is very difficult. If we can make use of the unlabeled examples when we train the learning machine, that means the data information is added and the classification accuracy will be improved.In this article, neighborhood search algorithm and semi-supervised algorithm are introduced to optimize the covering constructive learning method based on analyzing this method, besides, covering algorithm based on neighborhood search and semi-supervised covering algorithm are proposed. The main task is as follows:(1)This article summarizes the basic theory and algorithm of machine learning and artificial neural network. At the same time the model and patulous model of covering algorithm is introduced in detail and we also analyze the superiority and existent problem of the existing covering algorithm.(2)Neighborhood search algorithm is a simple and effective local search algorithm. Its basic criterion is using iterative method in the solution neighborhood to gradually optimize the objective function, until the date can not be optimized. In this article, neighborhood search algorithm and the covering constructive learning algorithm are combined and covering algorithm based on neighborhood search is proposed, which can optimize the covering network. In this way, every cover will contains more same labeled points and the number of covers will be smaller, thus the running time can be reduced. Experiment shows that the number of coverage is decreased greatly use the algorithm.(3) Semi-supervised learning is a hot research topic in the field of machine learning in recent years. It is a learning algorithm based on labeled samples and unlabeled samples. Human intervention is not needed throughout the learning process and the learning machine can uses the unlabeled examples itself. The core idea of semi-supervised learning is labeling the unlabeled samples, so as to increase the number of learning samples and improve the accuracy of machine learning. The superiority of semi-supervised learning is reflected in the simultaneous use of labeled samples and unlabeled samples. In this article, semi-supervised method is combined with the covering algorithm and a new algorithm named semi-supervised covering is proposed which can overcome some deficiencies of the covering constructive learning as a supervised learning. Experiment shows that this method can effectively improve recognition accuracy of the covering constructive learning and improve the generalization ability of the learning machine.