Learning Bayesian Network Structure With Priors And The Application To Intelligent Decision

As a kind of causal probability graph model,Bayesian Networks(BN)is a powerful tool to deal with the problem of uncertainty.In the use of BN to solve practical problems,we must first build BN structure,the BN structure can be constructed from expert experience or be learned from the data.At present,the common method for obtaining BN structure in practical problems is to learn the structure with priors,which can effectively combine the experts' priors with the sample information to obtain a more accurate model.Therefore,it is necessary to study the methods of learning BN structure with priors.For example,when BN is applied to the intelligent decision-making of Unmanned Aerial Vehicles(UAV),it is necessary to construct a reliable network structure,then how to deal with the incorrect prior knowledge of experts and how to conduct exact learning given certain priors remains to be studied.In view of the above problems,this dissertation carries out two aspects of studies including BN structure learning with priors and its application to UAV intelligent decision problem.The main research contents are as follows:(i)In this thesis,a new type of method of learning structure with uncertain priors is proposed.By studing the score-search method in depth,the method uses priors in the two processes of scoring and searching.First,a penalty term based priors scoring function with uncertain priors incorporating is proposed,in which the concept of entropy is introduced and how to make it decomposable is studied.Second,this thesis proposes a search strategy incorporating uncertain priors and enhance the robustness of priors utilizing,which is suitable for any heuristic search method.Finally,the hill-climbing method is used to verify the proposed method.The simulation results show that this method can utilize the correct prior information effectively and has strong adaptability to the wrong prior information.(ii)This thesis proposes a priors incorporating structure learning method based on dynamic programming.By studying the pattern and connotation of dynamic programming,we limit the process with the constraints of edges and paths,so that certain edge and path constraints are integrated into the process of dynamic programming BN structure learning.Specifically,constraints are used to restrict the planning process of the dynamic programming and to reduce the planning space.On the other hand,the sparse parent graph is further studied and the optimal parent selection is restricted by using constraints,to ensure that the optimal structure is in line with priors.Finally,the method based on priors incorporating is compared with the method without priors incorporating,the simulation results not only verify the effectiveness of the method in this thesis,but also prove that the incorporating of prior knowledge can obviously improve the learning efficiency and improve the learning accuracy.(iii)This thesis proposes a BN modeling method for UAV intelligent attack decision.This thesis studies UAV attack decision-making problem in detail,then designs the basic idea and procedure of Bayesian network modeling method for this problem.Flollowing the concrete steps of modeling are given: First,according to the assumption of UAV air-to-air combat,the basic mathematic model of attack decision-making is established,and then the related priors of structure is determined,next the methods proposed in this paper are used to study the structure,and then the parameters are learned.Finally,how to concretely use BN model for attack decision is illustrated and explained.