To Believe or Not To Believe: Improving Distributed Data Fusion with Second Order Knowledge

Most of works on Distributed Data Fusion (DDF ) systems investigate how to optimize or improve the fusion process by optimistically assuming the correctness of uncertainty models. The impact of using poor-quality information is not well addressed. Moreover, only a few works address how to improve the overall quality of the estimation result. The aim of this thesis is to cover this aspect of MultiAgent Systems and give an improvement on designing DDF frameworks, by using second order knowledge. In particular, we show that it is possible to improve the effectiveness of the estimation by using information not related only to the underlying filtering process. Understanding and reasoning about the context wherein the fusion process is performed is a key point for providing better and reliable estimations of the environment. Information about the intrinsic characteristics of the environment, the relations among fused data and the interactions among objects, all of them influence the performance of Data Fusion applications. The main result of this thesis is a complete framework for information source selection and integration, by using the reliability of sources as second order knowledge. Specifically, we developed a Multi-Agent Multi-Object Data Fusion system and applied it to the problem of Cooperative Robot Tag game and Soccer Robots. We address the question of reliability and existence of prior knowledge that impacts on the quality of data fusion process applied in such contexts. The chosen scenarios are well suited to test the effectiveness of our proposal: the dynamic component of these test-beds, as well as the competitiveness of the teams, define an ideal environment for demonstrating the results of this thesis. This notwithstanding, our system can be applied to many real-world applications: surveillance, people tracking, cluster maintenance, to name a few. The experimental results obtained show that the system outperforms the standard algorithms. Finally, we show how the system can be easily generalized to handle any kind of Multi-Sensor Tracking applications, such as Self-Tracking.