Research On Bayesian Network Structure Learning And Application

Bayesian Network(BN),as a probabilistic graph model combining probability theory and graph theory,has powerful knowledge representation and reasoning capability,and provides a solid theoretical foundation for solving problems such as prediction,classification,and reasoning.BN has been widely used in reliability and risk analysis,medical diagnosis,bioinformatics and other fields.When solving practical application problems based on BN theory,a key task is to establish a BN topology structure that can describe the relationship between different attributes according to the characteristics of the research object.Therefore,BN structure learning has gradually become a hot and difficult problem in this research field.This dissertation first studies the general BN structure learning algorithm,then optimizes the structure learning method of special BN(Bayesian network classifier,BNC),and finally applies BNC to the field of petroleum geology to solve the problem of spatial distribution of oil and gas resources.The main research work of the dissertation are as follows:(1)Since learning the optimal BN structure is an NP-hard problem,the swarm intelligence optimization methods are usually used to solve this problem,however these methods still have problems such as low search efficiency and optimization accuracy.To solve the above problems,this dissertation introduces firefly algorithm into BN structure learning for the first time,and proposes a novel method based on discrete firefly optimization algorithm to learn BN structure(DFA-B).DFA-B algorithm has the advantages of few parameters,fast convergence speed and strong global search ability.In DFA-B,firstly,the BN topological structure and the score of the corresponding structure are abstracted into the firefly's position and luminous intensity,respectively.Then,the flight strategy of fireflies in discrete states is redefined,and a mutation operator is added to each firefly for reference to the idea of evolutionary computing,which enhances its exploration ability and prevents the algorithm from falling into local optimum.Finally,a local optimizer is used to improve firefly's development capabilities,ensuring that the highest scoring candidate solutions are found.The algorithm proposed in this dissertation is compared with other algorithms on the benchmark networks.The experimental results show that the DFA-B has better score and better convergence.(2)TAN is a popular BNC method,but in the process of building the network structure,this method has two shortcomings:(1)randomly selecting the root node causes the fluctuation of classification performance,and(2)the existence of redundant edges between attributes will lead to the poor classification performance.To solve the above problems,this dissertation optimizes TAN structure from two aspects: the weighting between attributes and filtering redundant edges,and proposes a novel flexible tree-augmented na?ve bayes algorithm(FTAN).In FTAN,firstly,the contribution rate of mutual information(ICR)is defined to describes the relative amount of uncertainty reduction between different attributes.In the process of establishing the maximum weighted spanning tree,ICR is used to describe the dependencies between attributes in fine granularity,which is helpful to determine the direction of edges between attributes,so as to avoid the fluctuation of classification performance caused by random selection of root node.Then a flexible filtering method is adopted to filter out edges with weakly dependent relationships between attributes by dynamically adjusting the threshold.Experimental results reveal that the FTAN algorithm has significant advantages over other algorithms from the analysis of 0-1 loss function and class probability root mean square error.(3)Accurately predicting the spatial distribution of oil and gas resources is an important part of petroleum exploration,and it is of vital importance for improving exploration efficiency,optimizing drilling strategies,and increasing economic benefits.The existing methods to solve the problem of spatial distribution prediction of oil and gas resources are mainly modeling and prediction from a discriminative point of view.This type of method cannot quantitatively express the potential relationship between different geological attributes,and the classification accuracy has great room for improvement.The structure of BNC can naturally describe the relationship between different geological variables.Therefore,this dissertation applies BNC to the field of petroleum geology and proposes the BNC method to solve the problem of spatial distribution prediction of oil and gas.In the process of applied research,firstly,the mathematical expression of the prediction problem of oil and gas spatial distribution is defined.Then the KDB algorithm is optimized from the attributes sorting and model simplification,and the SKDB algorithm is proposed.Finally,two BNC methods,FTAN and SKDB,are used to solve the problem of oil and gas spatial distribution prediction of the Jurassic Sangonghe Formation in the hinterland of Junggar Basin.The application results show that the BNC method is significantly better than other currently popular methods in terms of accuracy and application effects.At the same time,according to the prediction results of the SKDB algorithm,it points out the favorable distribution area of the remaining oil and gas resources in the Sangonghe Formation,which provides an important reference for further exploration decision-making.