Research On Incremental Learning For Covering Algorithm

Recently, with the advancement of times and the rapid development of the information technology, hundreds of data continuously change over time in practical applications, and at the same time, it is very difficult and sometimes impractical to get a complete data set as training samples for machine learning algorithms at the beginning. Faced with these large-scale, continuously updating, dynamic and time-varying information data, if all the data made up of new data and original data every time are learned, it is not only resulting to huge waste of time and space, but also destroying the continuity of learning, and falling short of the human gradual learning habits. Therefore, there is an urgent need for researching effective machine learning algorithms and models to solve these issues.Incremental learning method is one of effective ways that learn the updating data continuously and deal with the classification problem of knowledge from massive data. It forms a continuous learning process by only learning new data every time, and retaining the existing learning outcomes. At present, there are some algorithms of the incremental learning, which are based on support vector machine, neural network, decision tree or the combine of decision tree and neural network and so on.The constructive machine learning method-covering algorithm possesses faster speed, lower complexity, stronger interpretability and higher precision. It solved the problem of the supervised learning effectively and achieved the favorable performance. In this dissertation, the covering algorithm and incremental learning are combined, and an incremental learning algorithm for the covering algorithm is proposed. Then considering the continuous occurrence of new samples and the problem of concept drift, three removing and forgetting mechanisms are introduced, and these forgetting mechanisms based on the covering algorithm are presented in detail. Moreover, the research on incremental learning for the covering algorithm provides an idea or a way for implementing classify in dynamic learning, it can also help us to solve many practical problems. The dissertation includes:1. An overview for the status of the incremental learning, including the research background, the significance, the current research in and around the world, as well as the classification methods of incremental learning that have been used are introduced at length.2. The covering algorithm and its improved algorithm are described. As the construction of the weight of the neurons--for new center of sphere domain is usually given to a man-made criteria, does not follow the distribution of samples to achieve the optimal solution. So we improve the covering algorithm inspired by Good-Point-Set.3. The covering algorithm and incremental learning are combined, and the incremental learning algorithm for the constructive covering method is proposed. This incremental learning algorithm for the constructive covering method uses the improved covering algorithm as base classifiers and refines the existing model through testing some new samples repeated, which reflects the "progressive" study of samples. The experimental results with the standard dataset show that the validity of the novel incremental learning algorithm.4. Three forgetting and removing algorithms are introduced, and these forgetting mechanisms based on the covering algorithm are explained. Finally, experiments have been carried out and analyzed.