DISTRIBUTED ADAPTATION AND LEARNING OVER COMPLEX NETWORKS

Abstract

Distributed information processing over networks has attracted much interest as the information is processed in a collaborative manner. In these systems, distributed agents are linked to each other and form an adaptive network, where each agent can communicate with its neighbors. With the help of this cooperation, the agents can solve particular tasks over the network. Among several strategies proposed to solve this task, the diffusion strategy is shown to be robust, stable, and capable of real-time adaptation. However, the literature on diffusion strategy has been mainly focused on single task networks, where all the agents have the same objective. This is an over-simplification for scenarios where there are various sources of intertwined information. In this thesis, a novel multitask diffusion strategy is proposed to solve the problem of distributed learning over multitask systems. Using the proposed method, two applications of distributed learning strategies are then tackled and evaluated. Lastly, the challenging problem of information reliability over networks is studied. An information credibility model is formulated and an adaptive reputation protocol is proposed to solve the problem.