Probabilistic modeling and reasoning in multiagent decision systems

Abstract

This thesis is about how to represent and solve multiagent decision problems in Bayesian decision theory. A new framework, including Multiply Sectioned Influence Diagrams (MSID) and Hyper Relevance Graph (HRG), is proposed based on influence diagrams. This new representation incorporates the idea of cooperative agentsa?? decision making and explicitly spells out the information support in agentsa?? decision making with respect to their organizational relationships. The theme of this thesis is to seek cooperative algorithms which coordinate the evaluation of local influence diagrams, and to seek efficient algorithms for solving an MSID. Thus three evaluation algorithms are proposed through extensions of basic evaluation approaches in influence diagrams and decision networks. Furthermore, a symbolic verification method is presented to develop a valid graphical decision model. In addition, the block learning algorithm is proposed to learn large Bayesian network structures from a small data set, which facilitates the construction of graphical decision models.