Iterative Message Alignment for Quantized Message Passing between Distributed Sensor Nodes

Mutual information maximizing clustering techniques, like the information bottleneck method, enable message passing based on compressed but highly informative beliefs. In this paper, we apply this concept to joint maximum a-posteriori detection problems in sensor networks. We show that by leveraging the information bottleneck method both the amount of exchanged data and the complexity of the operations performed in the involved sensor nodes respectively the fusion center is significantly reduced. In the considered network, distributed sensor nodes quantize their measurements and forward only cluster indices instead of high-precision cluster representatives to a fusion center. Due to a spatial distribution of the sensor nodes, the quantizers in the sensor nodes are optimized to the actual, varying measurement conditions. Thus, the meaning of a cluster index is sensor-dependent and cannot be uniquely recaptured if the transmitting sensor is unknown. Using a technique which we call message alignment we resolve this ambiguity without transmitting additional information to the fusion center. Additionally, we present a novel iterative message alignment algorithm to solve the generalized message alignment problem. Although only 4-bit integer-valued cluster indices are transmitted, included simulations show that our proposed system encounters no considerable performance degradation compared to an optimum maximum a-posteriori detection strategy.