The Research On Probabilistic Graphical Models For Uncertainty Processing In Distributed Computing Environments And Related Problems

Advances in computing and communication over networks have resulted in many perva-sive distributed computing environments, such as sensor networks, multi-agent systems andautonomous robot systems. These systems not only can distribute the computational tasksto different nodes, but also can (and are required to) collect information in different physicalspaces. In many circumstances, the enormous data they collect have contained many uncer-tainties, including noises, biases and uncertainties coming with imprecise or failed measur-ing. Reasoning under uncertainty to form coherent beliefs is one of the major tasks in thesedistributed systems. A major and prominent way to model and to process uncertainties isusing probabilistic graphical models.The"probabilistic graphical model"is an important research area in uncertainty rea-soning and artificial intelligence. There are a lot of models and algorithms in this area. Herethe problem is: how to use these models and algorithms for uncertainty processing in a dis-tributed environment. Further researches also include how to establish new graphical modelsand related algorithms for various distributed applications. Under this background problem,there are a bunch of smaller problems confronted: 1) For the data containing uncertaintiesthat are collected in a distributed way, how to build graphical models, including the parame-ters, according to dependencies among them. 2) For uncertainties in time sensitive data andsequential data, how to establish dynamic graphical models in the distributed environment?In more detailed analysis, that is how to solve problems such as the dynamic behaviors, theasynchronous behaviors, and the robustness of the system. 3) In many circumstances theparameters of some graphical models required in use have to be updated. How to establishupdate rules for parameter estimation in these models, and use them in a distributed envi- ronment? 4) Some graphical models applied to distributed systems can be context enabled.How to model the contextual effect on these models?This dissertation has studied the theories and related algorithms of probabilistic graphi-cal models, including Bayesian networks, Markov networks, factor graphs, standard dynamicBayesian networks, continuous time Bayesian networks, and some context enabled models.At the same time, this dissertation has investigated the characters and behaviors in distributedcomputing environments, such as in sensor networks. As a result, this dissertation makes thefollowing contributions in the related area:1. By investigating a general hybrid graphical model for uncertain data processing in sen-sor networks, we study the problem of how to learn parameters of this model in a dis-tributed way, and give a solution, including solution to parameter updating. Throughanalysis of this hybrid model, we find it has too strong limitations in distributed en-vironment usages for its undirected edges. So we build an extended model that onlyhas directed edges, with reasoning and parameter learning methods. It can avoid thelimitations, and be widely used.2. For modeling dynamic behaviors and processing time sequential data in distributedenvironments, we build a dynamic model based on standard dynamic Bayesian net-works. Comparing to the above static models, it can re?ect history effects on currentstates, which results to more precise modeling, with tracking the system dynamics. Itcan be easily realized. At the same time, we find that the synchronization problemmakes it limited in small scale networks. Then we present a new dynamic model usingcontinuous time Bayesian networks. In this model, we have solved the problems of dy-namic modeling in uneven time intervals; modeling asynchronous behaviors naturallyin distributed environments; the algorithm for belief updating and propagation; theway that different belief messages are organized in the system; the robustness of themodel when part of its host system is broken or severely delayed. It's a more complexmodel for distributed computing environments in large scale networks.3. We present update rules for parameter estimation in continuous time Bayesian net-works. It's an update learning algorithm. Comparing to other general parameter learn-ing algorithms for this state-of-art model, it can update the parameters from the oldmodel by new training data. We take analysis of the algorithm. We also use it forparameter updating of the above continuous time models in distributed environments.4. For modeling contextual effects on general graphical models, we present the FactorMetanetwork model. Comparing to Bayesian Metanetwork that only model directed relations, it can model both directed and undirected relations, with various expres-sions of parameter factors. We give cases of distribution computations under differentcontextual effects, and the general reasoning method for this model. We also give anexample application for sensor data fusion in different environmental circumstances.