Qualitative Probabilistic Network Based On Information Theory Conflict Reasoning

As a classic tool of uncertain knowledge representation and reasoning in artificial intelligence, Bayesian Network (BN) has been widely used in data mining, medical diagnosis, pattern recognition, industrial control, voice recognition, genetic chain analysis and so on. BNs can be used to discover the laws contained in the data, so it has attracted more and more attention. However, the construction and reasoning of BNs is too precise, resulting in that they cannot meet the requirement of effectiveness of many real-time applications.As the qualitative abstraction of BN, qualitative probabilistic network (QPN) simplifies uncertain knowledge representation and accelerates uncertain knowledge reasoning. QPN encodes the relationships among random variables by qualitative signs rather than conditional probabilities. Morever, QPN has an efficient reasoning algorithm. With the simplicity and efficiency, QPN can be appropriately used to remedy the above drawback of BN. However, the high abstraction level of QPN results in that inference conflicts always take place during QPN reasoning. Worst of all, when the ambiguous results caused by inference conflicts are produced, they will be spread to most part of the network with the reasoning algorithm going on. Therefore, inference conflicts have become a major application obstacle of QPN.In this thesis, in order to resolve QPN inference conflicts and remove QPN application obstacle at the same time, we first extend traditional QPN by adding a numerical weight to each qualitative influence of QPN so that inference conflicts can be avoided by comparison of the corresponding weights. Extended QPN is called MI-QPN. Then, we impement QPN conflict-free inference by extending the traditional reasioning algrithom based on MI-QPN.The main contribution of this work and can be summarized as follows:●We first define the mutual-information-based weight of a qualitative influence, called MIweight, and extend the traditional QPN by adding a Mlweight to each qualitative influence.●For both the QPN constructed directly from history data and the QPN abstracted from Bayesian network, we give two methods to derive MIweights of qualitative influences respectively.●We consequently discuss the symmetry, transitivity and composition properties of qualitative influences with MIweights in MI-QPN. Then, base on the three properties, we extend the general QPN's inference algorithm to achieve conflict-free inferences with the MI-QPN.●We finally make experiments to verify the feasibility of our methods. In addition, based the proposed theoretical method, we further design and develop a software prototype system, by which the progress of resolving QPN inference conflicts can be shown.