Distributed Boosting Variational Inference Algorithm Over Multi-Agent Networks

Distributed Bayesian estimation over multi-agent networks has received much attention due to its broad applications, where each agent has its private data that is unavailable to other agents. For efficient inference over multi-agent networks, we develop a distributed boosting variational inference (DBVI) algorithm with limited communication. We first decompose the global cost function into a
sum-of-costs
form, where each local cost only relates to its own dataset. Then, the global posterior distribution is approximated by a gradient decent at each boosting step, followed by a consensus protocol for cooperation with the neighbors. Moreover, we derive DBVI with Gaussian mixture model (DBVI-GMM) in detail. Finally, simulations on the synthetic and real datasets illustrate the effectiveness of the proposed algorithm.