Proceedings Article | 18 April 2006
KEYWORDS: Information fusion, Data fusion, Data analysis, Bayesian inference, Reconnaissance, Space reconnaissance, Information architecture, Criminalistics, Signals intelligence, Computer architecture
A new architecture for fusing information and data from heterogeneous sources is proposed. The approach takes criminalistics as a model. In analogy to the work of detectives, who attempt to investigate crimes, software agents are initiated that pursue clues and try to consolidate or to dismiss hypotheses. Like their human pendants, they can, if questions beyond their competences arise, consult expert agents. Within the context of a certain task, region, and time interval, specialized operations are applied to each relevant information source, e.g. IMINT, SIGINT, ACINT,..., HUMINT, data bases etc. in order to establish hit lists of first clues. Each clue is described by its pertaining facts, uncertainties, and dependencies in form of a local degree-of-belief (DoB) distribution in a Bayesian sense. For each clue an agent is initiated which cooperates with other agents and experts. Expert agents support to make use of different information sources. Consultations of experts, capable to access certain information sources, result in changes of the DoB of the pertaining clue. According to the significance of concentration of their DoB distribution clues are abandoned or pursued further to formulate task specific hypotheses. Communications between the agents serve to find out whether different clues belong to the same cause and thus can be put together. At the end of the investigation process, the different hypotheses are evaluated by a jury and a final report is created that constitutes the fusion result.
The approach proposed avoids calculating global DoB distributions by adopting a local Bayesian approximation and thus reduces the complexity of the exact problem essentially.
Different information sources are transformed into DoB distributions using the maximum entropy paradigm and considering known facts as constraints. Nominal, ordinal and cardinal quantities can be treated within this framework equally. The architecture is scalable by tailoring the number of agents according to the available computer resources, to the priority of tasks, and to the maximum duration of the fusion process. Furthermore, the architecture allows cooperative work of human and automated agents and experts, as long as not all subtasks can be accomplished automatically.
