Complexity problems in cross-layer optimization for wireless sensor networks have rekindled interest in the quest for a distributed algorithm that would simultaneously handle routing, task assignment, and information fusion, scale gracefully with the network size, and display resilience to node failure. Although various obstacles to this goal remain, belief propagation, and its close cousin expectation propagation, provide many attributes that such a distributed algorithm would require. Indeed, recent works attest to the potential of belief propagation for such tasks as network averaging, node detection, network diagnostics, routing, and even missile defense. This collaborative research project brings together researchers in statistical inference and wireless communications to (i) rephrase random sensor deployment strategies as a sparse dependency structure among parameters; (ii) advance expectation propagation as a distributed algorithm to harmonize many sensor network tasks; (iii) extend convergence results from iterative decoding to inference in sensor networks; and (iv) open novel design and optimization tools in sensor networks as by-products of the work.
In particular, the investigators show how common network inference tasks, including intruder detection, sensor localization, and channel estimation, can be viewed as particular instances of the expectation propagation algorithm passing messages between network nodes. Message passing here consists of soft information exchange, reminiscent of the mature field of iterative decoding. Convergence tools of iterative decoding, including density evolution and EXIT chart analysis, are extended to the network inference problems under consideration. Additional insights and new design tools for sensor networks emerge as natural by-products, ultimately targeting the inference capacity of such networks, and how this capacity may be optimized versus sleep strategies and energy consumption.
