DECENTRALIZED FUSION OF EXPERTS OVER NETWORKS

This article considers a network of agents interested in solving a classification t ask. The datasets available to accomplish the task are heterogeneous and dispersed across the agents. Each agent is inter-ested in discriminating among the 'inner' hypotheses reflected in its individual dataset. Moreover, the datasets at the different agents can be further labeled in terms of additional characteristics, giving rise to an enlarged space of hypotheses. The agents have no local information to distinguish their own dataset from the datasets of other agents. To overcome this issue, they want to share their individual knowledge to build an overall model that is able to address the classification task comprising all possible hypotheses. Starting from the optimal Bayesian fusion rule, we develop a strategy nicknamed de-centralized fusion of experts (DeFoE), which is able to build a global classifier starting from the classifiers locally available to the agents, at a reduced complexity, without retraining them from scratch. The effectiveness of the proposed strategy is shown over a benchmark dataset containing real images.