Finding Decomposable Models for Efficient Distributed Inference over Sensor Networks

Graphical models have been widely applied in distributed network computation problems such as inference in large-scale sensor networks. While belief propagation (BP) based on message passing is a powerful approach to solving such distributed inference problems, one major challenge, in the context of wireless sensor networks, is how to systematically address the trade-off between energy efficiency and inference performance. In this paper, we consider a distributed structure optimization problem and investigate the impacts of graphical model structure on energy consumption and inference performance. We first formulate the problem as a multi-objective constrained combinatorial optimization problem and prove its NP-hardness. Then, we propose an efficient distributed heuristic to solve the problem in polynomial time. Through extensive simulations, using both real-world sensor network data and synthetic data, we empirically evaluate our proposed graphical model structure optimization framework. The simulation results demonstrate that the graphical model constructed by the proposed framework can efficiently trades off the performance of the inference algorithm (measured by the mean squared error) with the energy consumed by the inference algorithm (measured by the energy used in communication). In addition, our proposed framework provides valuable insights for network designers on designing efficient model selection algorithms for distributed inference problems.