Rough Set Theory And Its Applied Research In Artificial Neural Networks

Rough set theory is a relatively new soft computing tool to deal with vagueness and uncertainty. It has received much attention of the researchers around the world. Rough set theory has been applied to many areas successfully including artificial intelligence, pattern recognition etc. This paper places emphasis on the study of the basic problems about rough set theory-knowledge reduction and real-value attributes discretization, discusses the generalized rough set models, and proposes three different methods in three different areas of combining rough set theory and neural network together based on these theories.This paper introduces the basic theory and conception of rough set in detail. And the main contribution of the paper in the framework of these as follows:1. The problem of knowledge reduction in rough set theory. Above all, this paper analyzes the classic methods-based on discernibility matrix. Then the significance of attributes in decision table is defined from the viewpoint of information; a heuristic algorithm based on information entropy for reduction of knowledge is proposed.2. The problem of real-value attributes discretization. This paper analyzes the original greedy algorithm and its improved algorithm, integrates the advantages of several former algorithms, proposes using mutual information to define the significance of discretization points in decision table, and integrates the conception of the core in discretization points to judging the significance of discretization points, then brings forwards a new improved method about greedy algorithm based on mutual information, solving the limitation of former algorithms effectively.3. Some generalization are needed to remedy the limitation of standard rough set models in practical application, here we introduce two generalized rough set models: fuzzy-rough set and variable precision rough set model. The theories of them have been expounded, and the attributes reduction algorithms of them have been studied.4. The methods of integration of rough set theory and neural network together have been studied. This paper using the superiority of rough set in dealing with imprecision and uncertainty, makes pretreatment to data samples, reduces the attributes, decreases the dimensions of samples, gets the approximation values, and obtains the decision rules after optimal reduction. Using these rules to map training samples of neural networks, constructing the numbers of latent layers and neural cells, makes the neural network more logical, reduces the training time of network, and improves the precision of training and the ability of generalization. According to different patterns of practical application, we bring forwards different methods of coupling and algorithms of training, and get satisfying results.