Research On Case Intelligent System Based On Rough Sets And Multilayer Feedforward Neural Networks

Learning is the most important embodiment of human intelligence, approaching to research and investigation on knowledge, has been the goal that mankind constantly engage for since ancient Greek. Analogy is an important cognitive model of human sense, and also a kind of inferential study strategy allowing people to process knowledge reasoning course wherever they have similar characters. Cases are the integrated representation of human sense, logics and creativity, and naturally become a common artifice when people process experiential decision-making. As a new reasoning technique to construct intelligent systems, case-based reasoning (CBR) performing in the computer, is the great achievement for the simulation of human analogy learning. No doubt it becomes a forthgoer in the research of human analogy sense, and a successful practitioner in these investigations. Study case intelligent system is helpful to imitate human thinking and achieve a simulator of human intelligence.The primary task for intelligent information processing is how to acquire knowledges, express them and drill rules for decision-making from huge amounts of data. Especially for the dealing with the real-world problems full of uncertainty, incomplete knowledge, Rough Sets (RS) theory and Artificial Neural Networks (ANN) technology have been emerging an amazing capacity. Centering round how to accomplish key CBR processes with these tremendous data, the dissertation therefore puts forward synthesis reasoning techniques mainly on the basis of RS and Multilayer Feed Forward Neural Networks (MFNN), and focuses on how to improve CBR system's precision with high efficiency, enlarge its flexibility besides robustness. The dissertation mainly deals with the following items.Firstly, cases are regarded as the foundation for knowledge representation in case intelligent system, and may be represented in semi-structured, unstructured model, or even in natural language text. The dissertation studies the feasibility of analogy from logics and reasoning structure for CBR process, and gives some potential perplexities about human's knowledge reasoning chains. To implement case intelligence, some aspects besides knowledge structure must be well managed. Those problems, such as how to construct suitable case base, organize cases and maintain them, are very important influencing factors to acquire the best problem-solving cases from the former quickly and efficiently, when they solve new problems in case retrieval, undoubtedly they are strongly interrelated with the system efficiency in problem-solving.Secondly, the dissertation studies the current main reseaching models on uncertain knowledge and analyzes their relationships from the view of knowledge chains, so that we can comprehend human sense and its problem-solving method from a higher level. In search of the models or methods to perform those theories fusion with methods, the dissertation investigates synthesis reasoning technology mainly combined with CBR and RS from many aspects. The findings indicate that effective synthesis reasoning contributes to an essential knowledge repository vital for incomplete knowledge discovery to make informed strategic decisions.According to the real problem-solving demands, synthesis reasoning can combine various reasoning principles and integrate many methods; consequently it can apply many kinds of knowledge from various levels, and enhance the system's efficiency of reasoning by means of adaptable knowledge granularity. Gradually it shall tend to set up a universal granular computing platform, which can deal with complex and fuzzy information more efficient, and provide reliable techniques in construction case intelligent systems, enlarging the system's ability of problem-solving, a framework of case-intelligent DSS is therefore figured out.Thirdly, the dissertation studies intelligent case retrieval techniques. After investigating the behavior of MFNN together with many kinds of existent algorithms for case retrieval based on MFNN, besides widely used back propagation algorithm (BP), simulated annealing algorithm and their ameliorated algorithms, it points out that weaknesses such as having lower speed and local extreme value, are inherent in those algorithms, thus cannot be conquered thoroughly. Compared with those former algorithms, radial basis function network (RBF), which recognizes and classifies samples dependence on their non-linear distance through projecting them to a multi dimensional space, is a good similarity detector. The subsequent research indicates that RBF network has so many advantages like fast learning and global convergence, then the dissertation puts forward a model for similar case retrieval based on RBF network.Fourthly, the dissertation investigates constructive neural networks together with their learning methods, and masters the learning course of neural networks from overall viewpoint. Any algorithm shall engage with certain characters of the neural network, which is selected as one of the purpose and should be taken care of. It is a significant leap to study constructive neural networks, which is capable of large-scale problem solving. The study begins with the MP neuron model from the view of geometry, and traces into covering algorithm (CA) from forward propagation algorithm (FP, compared with BP), the former algorithm of constructive neural networks. The sequent investigations indicate that CA has quite a number of characters, for example, clear system structure, feasible component for combination, running fast company with high recognizing ratio, etc., and it can be easily integrated and constructed. Then, the dissertation puts forward a model-case intelligent system-based on constructive neural networks and covering algorithm.Finally, the dissertation investigates case knowledge base maintenance and puts forward some criteria for this cycle. The main knowledge base of the CBR system is the case library, and its learning ability is continuously transacting adding new cases, just as people accumulate their knowledge. The knowledge implicit in case library is involved in CBR process, such as case representation, case matching and case adaptation, thus case base maintenance (CBM) becomes one of the key problem in CBR research. As an important branch of the CBR community, CBM has developed some kinds of methods, but they should be restricted to their running conditions, such as CBR system's scale, aging effect even the working area, and CBM methods must be adapted for any changes.Considering redundant cases or inconsistent knowledge possibly caused by the changing environment, the dissertation puts forward a CBM strategy based on Similarity Rough Set technique, which can achieve real-time maintenance, and implement a controllable threshold according to the selected granularity as a real-time monitor.Aiming to the efficiency of case retrieval in the irreducible case library, for that CBR systems running in interactive domains like e-commerce, online helpdesk, fault diagnosis etc., and can easily reach thousands of cases, the dissertation puts forward another CBM strategy based on covering algorithm and MFNN to achieve CBM from both sides: One is employing Alternative-Covering Algorithm to partition the case library to many Covering Domains and thus realizing the selective filtering; the other is using MFNN to deal with case retrieval within the large-scale case library. Our experimental results indicate that the integrated method, which is especially feasible for the processing of vast, multiclass and high dimensional data, can effectively guarantee the system's usability and enhance its capability.