| This dissertation provides a framework for the use of Bayesian Networks in large agent system for distributed interpretation and data fusion. Unlike previous approaches this work provides a framework that allows the local design of an agent's Bayesian network without knowledge of the overall agent hierarchy. This local design criterion is semantically predicated on each agent's functioning as part of a producer-consumer relation represented by a DAG called the agent graph. In this graph the agents are nodes, and there exists a directed edge between each agent that produces a variable and the agent(s) that consume it. A communication language is devised that involves the passing of probabilistic messages, containing the producer's belief in a shared variable, to the consumers along the edges in the agent graph. These messages may be observational (“I saw this event”) or interventional (“I caused this event”). Use of a message passing approach to facilitate global belief updating leads to the familiar “rumor” problem: some variables have redundant influence on calculations of belief in lower agents. Two solutions to the rumor problem are introduced. The first, known as the communication solution, involves expanded message passing and removes the redundancies at the communication level. The second, known as the model solution, constructs a second local Bayesian network that, when fed the redundantly contaminated beliefs in the consumed variables, returns the correct values of those variables. A trade-off criterion for determining which solution is superior given certain graph parameters is provided. These approaches require more advanced types of local belief updating than current propagation algorithms provide. In addition to the traditional calculation of full (single-variable) belief, multiple joint beliefs may also be required. Moreover, a device must also be provided to allow the integration of not just state restrictive observations on variables, but also of full distributions that come as either observations or interventions. Such a device and the algorithms to work on it are presented in this work. This solution, known as the factor tree, not only accomplishes these goals but also proves to be competitive with the best known general belief update algorithms. |
