DUCG:一种新的动态不确定因果知识的表达和推理方法(Ⅰ):离散、静态、证据确定和有向无环图情况  

A New Methodology to Deal with Dynamical Uncertain Causalities (Ⅰ):The Static Discrete DAG Case

作　　者：张勤[1] 

机构地区：[1]北京航空航天大学计算机学院,北京100191

出　　处：《计算机学报》2010年第4期625-651,共27页Chinese Journal of Computers

基　　金：国家自然科学基金(60643006)资助

摘　　要：贝叶斯网络(BN)是国际上流行的处理不确定因果知识的表达和推理模型.文中指出:即使所有变量均为二状态,BN中的子变量也有单赋值和多赋值之分.在单赋值情况下适用的知识的简洁表达和推理方法在多赋值情况下不适用.为克服BN的上述及其它缺陷,文中提出了DUCG(Dynamical Uncertainty Causality Graph)理论模型,以图形方式简洁表达任何情况下的不确定因果关系,并基于证据化简图形和展开事件,以得到所关注假设事件及其状态概率表达式.此外,DUCG允许知识表达不完备,使其超越了BN理论框架.一个入侵者报警系统被用来解释DUCG理论.Bayesian Network （BN） is a prominent model to deal with knowledge representation and inference in case of uncertain causalities.This paper discusses the essential difference between single-valued and multi-valued cases.It is pointed out that even a binary child variable can be either single-valued or multi-valued,while the existing compact representations and the corresponding inference algorithms applicable in single-valued cases cannot be simply applied in multi-valued cases.To overcome this problem and others,a new model named as DUCG （Dynamical Uncertain Causality Praph） is presented.By introducing a set of new concepts,DUCG is able to compactly and graphically represent complex conditional probability distributions （CPDs） in different modules,irrespective of whether the cases of the modules are single-valued or multi-valued.The simple connection among separately constructed modules composes a final DUCG.Once the evidence is observed,the first inference step is to simplify the DUCG regardless of queries by applying the 10 rules presented in this paper.The second step is to apply the event outspread algorithm presented in this paper to calculate the updated probabilities of the queries still in concern based on the simplified DUCG,regardless of whether the variables are singly or multiply connected.Sometimes,the qualitative solution can be found by only simplifying DUCG.Correspondingly,the accuracy of parameters is less important in DUCG.Moreover,DUCG enables people to represent the knowledge only in concern but not enough to represent CPDs.In other words,DUCG does not have to represent the joint probability distribution over a set of variables,although it is able to.This incompleteness of representation and flexiable conditional causalities represented in DUCG,in addition to that DUCG is able to deal with the directed cyclic graph （DCG） to be addressed in next paper,etc,makes DUCG beyound BN.The example of an alarm system detecting intruder illustrates the DUCG methodology.

关 键 词：智能系统 知识表达 概率推理 因果关系 不确定性 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]